ports/openmpi
ports/openmpi
1
1
Форкнуть 0
Код Релизы Активность
6f8e76df7e
openmpi/ompi/mca/bcol/basesmuma/bcol_basesmuma_lmsg_bcast.c

1879 строки
57 KiB
C
Исходник Обычный вид История

coll/ml: add support for blocking and non-blocking allreduce, reduce, and allgather. The new collectives provide a signifigant performance increase over tuned for small and medium messages. We are initially setting the priority lower than tuned until this has had some time to soak in the trunk. Please set coll_ml_priority to 90 for MTT runs. Credit for this work goes to Manjunath Gorentla Venkata (ORNL), Pavel Shamis (ORNL), and Nathan Hjelm (LANL). Commit details (for reference): Import ORNL's collectives for MPI_Allreduce, MPI_Reduce, and MPI_Allgather. We need to take the basesmuma header into account when calculating the ptpcoll small message thresholds. Add a define to bcol.h indicating the maximum header size so we can take the header into account while not making ptpcoll dependent on information from basesmuma. This resolves an issue with allreduce where ptpcoll overwrites the header of the next buffer in the basesmuma bank. Fix reduce and make a sequential collective launcher in coll_ml_inlines.h The root calculation for reduce was wrong for any root != 0. There are four possibilities for the root: - The root is not the current process but is in the current hierarchy. In this case the root is the index of the global root as specified in the root vector. - The root is not the current process and is not in the next level of the hierarchy. In this case 0 must be the local root since this process will never communicate with the real root. - The root is not the current process but will be in next level of the hierarchy. In this case the current process must be the root. - I am the root. The root is my index. Tested with IMB which rotates the root on every call to MPI_Reduce. Consider IMB the reproducer for the issue this commit solves. Make the bcast algorithm decision an enumerated variable Resolve various asset failures when destructing coll ml requests. Two issues: - Always reset the request to be invalid before returning it to the free list. This will avoid an asset in ompi_request_t's destructor. OMPI_REQUEST_FINI does this (and also releases the fortran handle index). - Never explicitly construct or destruct the superclass of an opal object. This screws up the class function tables and will cause either an assert failure or a segmentation fault when destructing coll ml requests. Cleanup allgather. I removed the duplicate non-blocking and blocking functions and modeled the cleanup after what I found in allreduce. Also cleaned up the code somewhat. Don't bother copying from the send to the recieve buffer in bcol_basesmuma_allreduce_intra_fanin_fanout if the pointers are the same. The eliminates a warning about memcpy and aliasing and avoids an unnecessary call to memcpy. Alwasy call CHECK_AND_RELEASE on memsync collectives. There was a call to OBJ_RELEASE on the collective communicator but because CHECK_AND_RECYLCE was never called there was not matching call to OBJ_RELEASE. This caused coll ml to leak communicators. Make allreduce use the sequential collective launcher in coll_ml_inlines.h Just launch the next collective in the component progress. I am a little unsure about this patch. There appears to be some sort of race between collectives that causes buffer exhaustion in some cases (IMB Allreduce is a reproducer). Changing progress to only launch the next bcol seems to resolve the issue but might not be the best fix. Note that I see little-no performance penalty for this change. Fix allreduce when there are extra sources. There was an issue with the buffer offset calculation when there are extra sources. In the case of extra sources == 1 the offset was set to buffer_size (just past the header of the next buffer). I adjusted the buffer size to take into accoun the maximum header size (see the earlier commit that added this) and simplified the offset calculation. Make reduce/allreduce non-blocking. This is required for MPI_Comm_idup to work correctly. This has been tested with various layouts using the ibm testsuite and imb and appears to have the same performance as the old blocking version. Fix allgather for non-contiguous layouts and simplify parsing the topology. Some things in this patch: - There were several comments to the effect that level 0 of the hierarchy MUST contain all of the ranks. At least one function made this assumption but it was not true. I changed the sbgp components and the coll ml initization code to enforce this requirement. - Ensure that hierarchy level 0 has the ranks in the correct scatter gather order. This removes the need for a separate sort list and fixes the offset calculation for allgather. - There were several passes over the hierarchy to determine properties of the hierarchy. I eliminated these extra passes and the memory allocation associated with them and calculate the tree properties on the fly. The same DFS recursion also handles the re-order of level 0. All these changes have been verified with MPI_Allreduce, MPI_Reduce, and MPI_Allgather. All functions now pass all IBM/Open MPI, and IMB tests. coll/ml: correct pointer usage for MPI_BOTTOM Since contiguous datatypes are copied via memcpy (bypassing the convertor) we need to adjust for the lb of the datatype. This corrects problems found testing code that uses MPI_BOTTOM (NULL) as the send pointer. Add fallback collectives for allreduce and reduce. cmr=v1.7.5:reviewer=pasha This commit was SVN r30363.
2014-01-22 19:39:19 +04:00
/*
* Copyright (c) 2009-2012 Oak Ridge National Laboratory. All rights reserved.
* Copyright (c) 2009-2012 Mellanox Technologies. All rights reserved.
* $COPYRIGHT$
*
* Additional copyrights may follow
*
* $HEADER$
*/
#include "ompi_config.h"
#ifdef __PORTALS_AVAIL__
#define __PORTALS_ENABLE__
#include "ompi/constants.h"
#include "ompi/datatype/ompi_datatype.h"
#include "ompi/communicator/communicator.h"
#include "bcol_basesmuma.h"
#include "bcol_basesmuma_portals.h"
#include "bcol_basesmuma_lmsg_bcast.h"
#include "bcol_basesmuma_utils.h"
/*
* Scatter/Gather Broadcast algorithm
*
* Algorithm highlights:
*
* Uses portals for data transfer
*
* All processes participating in the broadcast are arranged in a
* binmoial tree.
*
* Phase1: Scatter the broadcast data to all the children
* Phase2: All processes in the tree participates in recursive doubling
* algorithm to obtain the missing data.
*/
static int completed_scatter = 0;
#if 0
int bcol_basesmuma_lmsg_scatter_allgather_portals_bcast_old(bcol_function_args_t *input_args,
coll_ml_function_t *c_input_args)
{
/* local variables */
int i;
uint64_t length;
int my_rank, parent_rank, src =-1, matched = 0;
int *src_list = NULL;
int group_size = -1, dummy_group_size;
int first_instance=0;
int rc = OMPI_SUCCESS;
int leading_dim, buff_idx, idx;
int count=input_args->count;
size_t pack_len = 0, dt_size =0 ;
int64_t ready_flag;
int flag_offset;
int pow_2, pow_2_levels;
int src_list_index = -1;
uint64_t fragment_size; /* user buffer size */
int sg_matchbits = 0;
/* Input argument variables */
void *my_userbuf = (void*)((unsigned char*)input_args->userbuf);
int64_t sequence_number=input_args->sequence_num;
struct ompi_datatype_t* dtype=input_args->dtype;
/* Extra source variables */
bool secondary_root = false;
int partner = -1, extra_partner = -1;
/* Scatter Allgather offsets */
uint64_t local_sg_offset = 0, global_sg_offset = 0, partner_offset = 0;
/* Portals messaging relevant variables */
mca_bcol_basesmuma_portal_proc_info_t *portals_info;
ptl_handle_eq_t allgather_eq_h;
ptl_event_t allgather_event;
bool blocked_post = false;
bool msg_posted = false;
int total_msg_posts = -1, scatter_posts = -1, allgather_posts = -1, extra_src_posts = -1;
/* OMPI module and component variables */
mca_bcol_basesmuma_component_t *cs = &mca_bcol_basesmuma_component;
mca_bcol_basesmuma_module_t *bcol_module =
(mca_bcol_basesmuma_module_t *) c_input_args->bcol_module;
/* Control structure and payload variables */
volatile mca_bcol_basesmuma_payload_t *data_buffs;
volatile mca_bcol_basesmuma_header_t *my_ctl_pointer = NULL;
volatile mca_bcol_basesmuma_header_t *parent_ctl_pointer = NULL;
volatile mca_bcol_basesmuma_header_t *partner_ctl_pointer = NULL;
struct mca_bcol_basesmuma_portal_buf_addr_t *my_lmsg_ctl_pointer = NULL;
struct mca_bcol_basesmuma_portal_buf_addr_t *parent_lmsg_ctl_pointer = NULL;
struct mca_bcol_basesmuma_portal_buf_addr_t *partner_lmsg_ctl_pointer = NULL;
/* Make sure there userbuffer is not null */
assert(my_userbuf != NULL);
/* Get portals info*/
portals_info = (mca_bcol_basesmuma_portal_proc_info_t*)cs->portals_info;
/* Get addresing information */
buff_idx = input_args->src_desc->buffer_index;
group_size = bcol_module->colls_no_user_data.size_of_group;
leading_dim=bcol_module->colls_no_user_data.size_of_group;
my_rank = bcol_module->super.sbgp_partner_module->my_index;
idx=SM_ARRAY_INDEX(leading_dim,buff_idx,0);
/* calculate the largest power of two that is smaller than
* or equal to the group size
*/
pow_2_levels = pow_sm_k(2,group_size, &(dummy_group_size));
if( group_size < (1<<pow_2_levels)) {
pow_2_levels--;
}
/* power-of-two group size */
pow_2 = 1<<pow_2_levels;
/* we will work only on packed data - so compute the length*/
ompi_datatype_type_size(dtype, &dt_size);
fragment_size = count*dt_size;
/* grab the data buffs */
data_buffs = (mca_bcol_basesmuma_payload_t *)
bcol_module->colls_with_user_data.data_buffs+idx;
my_ctl_pointer = data_buffs[my_rank].ctl_struct;
my_lmsg_ctl_pointer = (mca_bcol_basesmuma_portal_buf_addr_t*)
data_buffs[my_rank].payload;
if(my_ctl_pointer->sequence_number < sequence_number) {
first_instance = 1;
}
if(first_instance) {
my_ctl_pointer->flag = -1;
my_ctl_pointer->index = 1;
my_ctl_pointer->starting_flag_value = 0;
flag_offset = 0;
} else {
my_ctl_pointer->index++;
}
assert( -1 == my_ctl_pointer->flag);
/* increment the starting flag by one and return */
flag_offset = my_ctl_pointer->starting_flag_value;
ready_flag = flag_offset + sequence_number + 1;
my_ctl_pointer->sequence_number = sequence_number;
sg_matchbits = sequence_number ;
/* Construct my portal buffer address and copy to payload buffer */
mca_bcol_basesmuma_construct_portal_address(my_lmsg_ctl_pointer,
portals_info->portal_id.nid,
portals_info->portal_id.pid,
sg_matchbits,
bcol_module->super.sbgp_partner_module->group_comm->c_contextid);
my_lmsg_ctl_pointer->userbuf = my_userbuf;
my_lmsg_ctl_pointer->userbuf_length = fragment_size;
/*
* If I am the root of bcast, scatter the data to my children
*/
if (input_args->root_flag) {
BASESMUMA_VERBOSE(10,("I am the root of the data"));
my_lmsg_ctl_pointer->offset = 0;
my_lmsg_ctl_pointer->n_sends = pow_2_levels;
my_lmsg_ctl_pointer->length = fragment_size;
rc = PtlEQAlloc(((mca_bcol_basesmuma_portal_proc_info_t*)
cs->portals_info)->ni_h, MAX_PORTAL_EVENTS_IN_Q,
PTL_EQ_HANDLER_NONE, &allgather_eq_h);
if (rc != PTL_OK) {
BASESMUMA_VERBOSE(10,( "PtlEQAlloc() failed: %d \n",rc));
goto Release;
}
/* Compute number of posts required
* We post the data buffer for both scatter and allgather phase at once so to avoid
* posting overhead
*/
if (my_rank >= pow_2) {
/* I am root and my rank is greater than pow_2, I will hand
* over to rank (that is < pow_2) to act as secondary root
*/
total_msg_posts = 1;
}
else {
extra_src_posts = (my_rank + pow_2 < group_size ) ? 1: 0;
scatter_posts = my_lmsg_ctl_pointer->n_sends;
allgather_posts = pow_2_levels - 1;
total_msg_posts = scatter_posts + allgather_posts + extra_src_posts ;
}
mca_bcol_basesmuma_portals_post_msg(cs, my_lmsg_ctl_pointer,
my_userbuf, fragment_size, allgather_eq_h,
total_msg_posts,
blocked_post,
PTL_MD_EVENT_START_DISABLE| PTL_MD_EVENT_END_DISABLE | PTL_MD_OP_GET | PTL_MD_MANAGE_REMOTE | PTL_MD_TRUNCATE | PTL_MD_EVENT_AUTO_UNLINK_ENABLE);
msg_posted = true ;
/* important that these be set before my children
* see the ready flag raised
*/
basesmuma: do not call MB directly opal does not always define MB. It is recommended that opal_atomic_[rw]mb is called instead. We will need to address the cases where these functions are no-ops on weak-memory ordered cpus. cmr=v1.7.5:ticket=trac:4158 This commit was SVN r30534. The following Trac tickets were found above: Ticket 4158 --> https://svn.open-mpi.org/trac/ompi/ticket/4158
2014-02-03 21:01:57 +04:00
opal_atomic_wmb ();
coll/ml: add support for blocking and non-blocking allreduce, reduce, and allgather. The new collectives provide a signifigant performance increase over tuned for small and medium messages. We are initially setting the priority lower than tuned until this has had some time to soak in the trunk. Please set coll_ml_priority to 90 for MTT runs. Credit for this work goes to Manjunath Gorentla Venkata (ORNL), Pavel Shamis (ORNL), and Nathan Hjelm (LANL). Commit details (for reference): Import ORNL's collectives for MPI_Allreduce, MPI_Reduce, and MPI_Allgather. We need to take the basesmuma header into account when calculating the ptpcoll small message thresholds. Add a define to bcol.h indicating the maximum header size so we can take the header into account while not making ptpcoll dependent on information from basesmuma. This resolves an issue with allreduce where ptpcoll overwrites the header of the next buffer in the basesmuma bank. Fix reduce and make a sequential collective launcher in coll_ml_inlines.h The root calculation for reduce was wrong for any root != 0. There are four possibilities for the root: - The root is not the current process but is in the current hierarchy. In this case the root is the index of the global root as specified in the root vector. - The root is not the current process and is not in the next level of the hierarchy. In this case 0 must be the local root since this process will never communicate with the real root. - The root is not the current process but will be in next level of the hierarchy. In this case the current process must be the root. - I am the root. The root is my index. Tested with IMB which rotates the root on every call to MPI_Reduce. Consider IMB the reproducer for the issue this commit solves. Make the bcast algorithm decision an enumerated variable Resolve various asset failures when destructing coll ml requests. Two issues: - Always reset the request to be invalid before returning it to the free list. This will avoid an asset in ompi_request_t's destructor. OMPI_REQUEST_FINI does this (and also releases the fortran handle index). - Never explicitly construct or destruct the superclass of an opal object. This screws up the class function tables and will cause either an assert failure or a segmentation fault when destructing coll ml requests. Cleanup allgather. I removed the duplicate non-blocking and blocking functions and modeled the cleanup after what I found in allreduce. Also cleaned up the code somewhat. Don't bother copying from the send to the recieve buffer in bcol_basesmuma_allreduce_intra_fanin_fanout if the pointers are the same. The eliminates a warning about memcpy and aliasing and avoids an unnecessary call to memcpy. Alwasy call CHECK_AND_RELEASE on memsync collectives. There was a call to OBJ_RELEASE on the collective communicator but because CHECK_AND_RECYLCE was never called there was not matching call to OBJ_RELEASE. This caused coll ml to leak communicators. Make allreduce use the sequential collective launcher in coll_ml_inlines.h Just launch the next collective in the component progress. I am a little unsure about this patch. There appears to be some sort of race between collectives that causes buffer exhaustion in some cases (IMB Allreduce is a reproducer). Changing progress to only launch the next bcol seems to resolve the issue but might not be the best fix. Note that I see little-no performance penalty for this change. Fix allreduce when there are extra sources. There was an issue with the buffer offset calculation when there are extra sources. In the case of extra sources == 1 the offset was set to buffer_size (just past the header of the next buffer). I adjusted the buffer size to take into accoun the maximum header size (see the earlier commit that added this) and simplified the offset calculation. Make reduce/allreduce non-blocking. This is required for MPI_Comm_idup to work correctly. This has been tested with various layouts using the ibm testsuite and imb and appears to have the same performance as the old blocking version. Fix allgather for non-contiguous layouts and simplify parsing the topology. Some things in this patch: - There were several comments to the effect that level 0 of the hierarchy MUST contain all of the ranks. At least one function made this assumption but it was not true. I changed the sbgp components and the coll ml initization code to enforce this requirement. - Ensure that hierarchy level 0 has the ranks in the correct scatter gather order. This removes the need for a separate sort list and fixes the offset calculation for allgather. - There were several passes over the hierarchy to determine properties of the hierarchy. I eliminated these extra passes and the memory allocation associated with them and calculate the tree properties on the fly. The same DFS recursion also handles the re-order of level 0. All these changes have been verified with MPI_Allreduce, MPI_Reduce, and MPI_Allgather. All functions now pass all IBM/Open MPI, and IMB tests. coll/ml: correct pointer usage for MPI_BOTTOM Since contiguous datatypes are copied via memcpy (bypassing the convertor) we need to adjust for the lb of the datatype. This corrects problems found testing code that uses MPI_BOTTOM (NULL) as the send pointer. Add fallback collectives for allreduce and reduce. cmr=v1.7.5:reviewer=pasha This commit was SVN r30363.
2014-01-22 19:39:19 +04:00
my_ctl_pointer->flag = ready_flag;
/* Wait for my scatter partner */
if (my_rank >= pow_2) {
int scatter_partner = -1;
volatile mca_bcol_basesmuma_header_t *scatter_partner_ctl_pointer = NULL;
scatter_partner = my_rank - pow_2;
scatter_partner_ctl_pointer =
data_buffs[scatter_partner].ctl_struct;
while(!IS_SG_DATA_READY(scatter_partner_ctl_pointer, ready_flag,
sequence_number)){
opal_progress();
}
goto Release;
}
else {
wait_for_peers(my_rank, my_lmsg_ctl_pointer->n_sends, data_buffs,
ready_flag, sequence_number);
}
goto Allgather;
}
Extra :
if( my_rank >= pow_2 ) {
parent_rank = my_rank & (pow_2-1);
parent_ctl_pointer = data_buffs[parent_rank].ctl_struct;
parent_lmsg_ctl_pointer =
(mca_bcol_basesmuma_portal_buf_addr_t*)data_buffs[parent_rank].payload;
ready_flag = ready_flag + pow_2_levels;
while(!IS_SG_DATA_READY(parent_ctl_pointer, ready_flag, sequence_number)) {
opal_progress();
}
mca_bcol_basesmuma_portals_get_msg_fragment_no_eq_h(cs, my_lmsg_ctl_pointer,
parent_lmsg_ctl_pointer, 0,
0, fragment_size);
my_ctl_pointer->flag = ready_flag;
goto Release;
}
Scatter:
/* I am not root of bcast compute the list of possible
* where I will receive bcast data from.
*/
src_list = (int *) malloc(sizeof(int) * (pow_2_levels + 1));
for( i = 0; i < pow_2_levels; i++) {
src_list[i] = my_rank ^ (1<<i);
}
/* My source might be process > pow_2 */
if ((my_rank + pow_2) < group_size) {
src_list[i] = my_rank + pow_2;
} else {
src_list[i] = -1;
}
Probe:
/* If I am not the root, then poll on possible "senders'" control structs */
/* For portals we block for now */
while (!matched) {
/* Shared memory iprobe */
SG_LARGE_MSG_PROBE(src_list, pow_2_levels + 1,
src_list_index, matched, src, data_buffs, parent_ctl_pointer,
parent_lmsg_ctl_pointer,ready_flag, sequence_number);
}
/* If I am a secondary root
* Secondary root acts as root of bcast data when real root of data
* is process with group rank greater than pow_2 */
if ((matched) && (src == pow_2 + my_rank)) {
volatile mca_bcol_basesmuma_header_t *extra_src_ctl_pointer = NULL;
struct mca_bcol_basesmuma_portal_buf_addr_t *extra_src_lmsg_ctl_pointer = NULL;
secondary_root = true;
BASESMUMA_VERBOSE(10,("I am the secondary root for the data"));
my_lmsg_ctl_pointer->offset = 0;
my_lmsg_ctl_pointer->n_sends = pow_2_levels;
my_lmsg_ctl_pointer->length = fragment_size;
extra_src_ctl_pointer = data_buffs[src].ctl_struct;
extra_src_lmsg_ctl_pointer = (mca_bcol_basesmuma_portal_buf_addr_t*)data_buffs[src].payload;
/* create an event queue for the incoming buffer */
rc = PtlEQAlloc(((mca_bcol_basesmuma_portal_proc_info_t*)
cs->portals_info)->ni_h, MAX_PORTAL_EVENTS_IN_Q,
PTL_EQ_HANDLER_NONE, &allgather_eq_h);
if (rc != PTL_OK) {
BASESMUMA_VERBOSE(10,( "PtlEQAlloc() failed: %d \n",rc));
goto Release;
}
mca_bcol_basesmuma_portals_get_msg_fragment_no_eq_h(cs, my_lmsg_ctl_pointer,
extra_src_lmsg_ctl_pointer, 0,
0, fragment_size);
extra_src_posts = 0;
scatter_posts = my_lmsg_ctl_pointer->n_sends;
allgather_posts = pow_2_levels - 1;
total_msg_posts = scatter_posts + allgather_posts + extra_src_posts ;
mca_bcol_basesmuma_portals_post_msg(cs, my_lmsg_ctl_pointer,
my_userbuf, fragment_size, allgather_eq_h,
total_msg_posts,
blocked_post,
PTL_MD_EVENT_START_DISABLE| PTL_MD_EVENT_END_DISABLE | PTL_MD_OP_GET | PTL_MD_MANAGE_REMOTE | PTL_MD_TRUNCATE | PTL_MD_EVENT_AUTO_UNLINK_ENABLE);
msg_posted = true ;
/* important that these be set before my children
* see the ready flag raised
*/
basesmuma: do not call MB directly opal does not always define MB. It is recommended that opal_atomic_[rw]mb is called instead. We will need to address the cases where these functions are no-ops on weak-memory ordered cpus. cmr=v1.7.5:ticket=trac:4158 This commit was SVN r30534. The following Trac tickets were found above: Ticket 4158 --> https://svn.open-mpi.org/trac/ompi/ticket/4158
2014-02-03 21:01:57 +04:00
opal_atomic_wmb ();
coll/ml: add support for blocking and non-blocking allreduce, reduce, and allgather. The new collectives provide a signifigant performance increase over tuned for small and medium messages. We are initially setting the priority lower than tuned until this has had some time to soak in the trunk. Please set coll_ml_priority to 90 for MTT runs. Credit for this work goes to Manjunath Gorentla Venkata (ORNL), Pavel Shamis (ORNL), and Nathan Hjelm (LANL). Commit details (for reference): Import ORNL's collectives for MPI_Allreduce, MPI_Reduce, and MPI_Allgather. We need to take the basesmuma header into account when calculating the ptpcoll small message thresholds. Add a define to bcol.h indicating the maximum header size so we can take the header into account while not making ptpcoll dependent on information from basesmuma. This resolves an issue with allreduce where ptpcoll overwrites the header of the next buffer in the basesmuma bank. Fix reduce and make a sequential collective launcher in coll_ml_inlines.h The root calculation for reduce was wrong for any root != 0. There are four possibilities for the root: - The root is not the current process but is in the current hierarchy. In this case the root is the index of the global root as specified in the root vector. - The root is not the current process and is not in the next level of the hierarchy. In this case 0 must be the local root since this process will never communicate with the real root. - The root is not the current process but will be in next level of the hierarchy. In this case the current process must be the root. - I am the root. The root is my index. Tested with IMB which rotates the root on every call to MPI_Reduce. Consider IMB the reproducer for the issue this commit solves. Make the bcast algorithm decision an enumerated variable Resolve various asset failures when destructing coll ml requests. Two issues: - Always reset the request to be invalid before returning it to the free list. This will avoid an asset in ompi_request_t's destructor. OMPI_REQUEST_FINI does this (and also releases the fortran handle index). - Never explicitly construct or destruct the superclass of an opal object. This screws up the class function tables and will cause either an assert failure or a segmentation fault when destructing coll ml requests. Cleanup allgather. I removed the duplicate non-blocking and blocking functions and modeled the cleanup after what I found in allreduce. Also cleaned up the code somewhat. Don't bother copying from the send to the recieve buffer in bcol_basesmuma_allreduce_intra_fanin_fanout if the pointers are the same. The eliminates a warning about memcpy and aliasing and avoids an unnecessary call to memcpy. Alwasy call CHECK_AND_RELEASE on memsync collectives. There was a call to OBJ_RELEASE on the collective communicator but because CHECK_AND_RECYLCE was never called there was not matching call to OBJ_RELEASE. This caused coll ml to leak communicators. Make allreduce use the sequential collective launcher in coll_ml_inlines.h Just launch the next collective in the component progress. I am a little unsure about this patch. There appears to be some sort of race between collectives that causes buffer exhaustion in some cases (IMB Allreduce is a reproducer). Changing progress to only launch the next bcol seems to resolve the issue but might not be the best fix. Note that I see little-no performance penalty for this change. Fix allreduce when there are extra sources. There was an issue with the buffer offset calculation when there are extra sources. In the case of extra sources == 1 the offset was set to buffer_size (just past the header of the next buffer). I adjusted the buffer size to take into accoun the maximum header size (see the earlier commit that added this) and simplified the offset calculation. Make reduce/allreduce non-blocking. This is required for MPI_Comm_idup to work correctly. This has been tested with various layouts using the ibm testsuite and imb and appears to have the same performance as the old blocking version. Fix allgather for non-contiguous layouts and simplify parsing the topology. Some things in this patch: - There were several comments to the effect that level 0 of the hierarchy MUST contain all of the ranks. At least one function made this assumption but it was not true. I changed the sbgp components and the coll ml initization code to enforce this requirement. - Ensure that hierarchy level 0 has the ranks in the correct scatter gather order. This removes the need for a separate sort list and fixes the offset calculation for allgather. - There were several passes over the hierarchy to determine properties of the hierarchy. I eliminated these extra passes and the memory allocation associated with them and calculate the tree properties on the fly. The same DFS recursion also handles the re-order of level 0. All these changes have been verified with MPI_Allreduce, MPI_Reduce, and MPI_Allgather. All functions now pass all IBM/Open MPI, and IMB tests. coll/ml: correct pointer usage for MPI_BOTTOM Since contiguous datatypes are copied via memcpy (bypassing the convertor) we need to adjust for the lb of the datatype. This corrects problems found testing code that uses MPI_BOTTOM (NULL) as the send pointer. Add fallback collectives for allreduce and reduce. cmr=v1.7.5:reviewer=pasha This commit was SVN r30363.
2014-01-22 19:39:19 +04:00
my_ctl_pointer->flag = ready_flag;
wait_for_peers(my_rank, my_lmsg_ctl_pointer->n_sends, data_buffs,
ready_flag, sequence_number);
goto Allgather;
}
/* Verify whether we got the right
* source of the data, by computing the source's intended
* destinations
*/
for( i = 0; i < parent_lmsg_ctl_pointer->n_sends; i++) {
uint64_t local_offset = 0;
uint64_t remote_offset = 0;
BASESMUMA_VERBOSE(5,("%d found it from %d \n",my_rank,src));
if( my_rank == (src^(1<<i))) {
parent_ctl_pointer = data_buffs[src].ctl_struct;
parent_lmsg_ctl_pointer = (mca_bcol_basesmuma_portal_buf_addr_t*)data_buffs[src].payload;
/* we found our root within the group ... */
BASESMUMA_VERBOSE(10,("Shared memory probe was matched, the root is %d", src));
my_lmsg_ctl_pointer->n_sends = i;
/* Am I source for other process during scatter phase */
if ( i > 0) {
/* compute the size of the chunk to copy */
length = (parent_lmsg_ctl_pointer->length)/
(1<<(parent_lmsg_ctl_pointer->n_sends - my_lmsg_ctl_pointer->n_sends));
my_lmsg_ctl_pointer->length = length;
my_lmsg_ctl_pointer->offset =
parent_lmsg_ctl_pointer->offset + length;
local_offset = my_lmsg_ctl_pointer->offset;
remote_offset = parent_lmsg_ctl_pointer->offset + length;
mca_bcol_basesmuma_portals_get_msg_fragment_no_eq_h(cs, my_lmsg_ctl_pointer,
parent_lmsg_ctl_pointer,local_offset,
remote_offset, length);
rc = PtlEQAlloc(((mca_bcol_basesmuma_portal_proc_info_t*)
cs->portals_info)->ni_h, MAX_PORTAL_EVENTS_IN_Q,
PTL_EQ_HANDLER_NONE,
&allgather_eq_h);
if (rc != PTL_OK) {
BASESMUMA_VERBOSE(10,( "PtlEQAlloc() failed: %d \n",rc));
goto Release;
}
/* Now post the message for other children to read */
extra_src_posts = (my_rank + pow_2 < group_size ) ? 1: 0;
scatter_posts = my_lmsg_ctl_pointer->n_sends;
allgather_posts = pow_2_levels - 1;
total_msg_posts = scatter_posts + allgather_posts + extra_src_posts ;
mca_bcol_basesmuma_portals_post_msg(cs, my_lmsg_ctl_pointer,
my_userbuf, my_lmsg_ctl_pointer->userbuf_length,
allgather_eq_h,
total_msg_posts,
blocked_post,
PTL_MD_EVENT_START_DISABLE| PTL_MD_EVENT_END_DISABLE
| PTL_MD_OP_GET | PTL_MD_MANAGE_REMOTE | PTL_MD_TRUNCATE | PTL_MD_EVENT_AUTO_UNLINK_ENABLE
);
msg_posted = true;
/* set the memory barrier to ensure completion
* and signal I am done getting scatter data*/
basesmuma: do not call MB directly opal does not always define MB. It is recommended that opal_atomic_[rw]mb is called instead. We will need to address the cases where these functions are no-ops on weak-memory ordered cpus. cmr=v1.7.5:ticket=trac:4158 This commit was SVN r30534. The following Trac tickets were found above: Ticket 4158 --> https://svn.open-mpi.org/trac/ompi/ticket/4158
2014-02-03 21:01:57 +04:00
opal_atomic_wmb ();
coll/ml: add support for blocking and non-blocking allreduce, reduce, and allgather. The new collectives provide a signifigant performance increase over tuned for small and medium messages. We are initially setting the priority lower than tuned until this has had some time to soak in the trunk. Please set coll_ml_priority to 90 for MTT runs. Credit for this work goes to Manjunath Gorentla Venkata (ORNL), Pavel Shamis (ORNL), and Nathan Hjelm (LANL). Commit details (for reference): Import ORNL's collectives for MPI_Allreduce, MPI_Reduce, and MPI_Allgather. We need to take the basesmuma header into account when calculating the ptpcoll small message thresholds. Add a define to bcol.h indicating the maximum header size so we can take the header into account while not making ptpcoll dependent on information from basesmuma. This resolves an issue with allreduce where ptpcoll overwrites the header of the next buffer in the basesmuma bank. Fix reduce and make a sequential collective launcher in coll_ml_inlines.h The root calculation for reduce was wrong for any root != 0. There are four possibilities for the root: - The root is not the current process but is in the current hierarchy. In this case the root is the index of the global root as specified in the root vector. - The root is not the current process and is not in the next level of the hierarchy. In this case 0 must be the local root since this process will never communicate with the real root. - The root is not the current process but will be in next level of the hierarchy. In this case the current process must be the root. - I am the root. The root is my index. Tested with IMB which rotates the root on every call to MPI_Reduce. Consider IMB the reproducer for the issue this commit solves. Make the bcast algorithm decision an enumerated variable Resolve various asset failures when destructing coll ml requests. Two issues: - Always reset the request to be invalid before returning it to the free list. This will avoid an asset in ompi_request_t's destructor. OMPI_REQUEST_FINI does this (and also releases the fortran handle index). - Never explicitly construct or destruct the superclass of an opal object. This screws up the class function tables and will cause either an assert failure or a segmentation fault when destructing coll ml requests. Cleanup allgather. I removed the duplicate non-blocking and blocking functions and modeled the cleanup after what I found in allreduce. Also cleaned up the code somewhat. Don't bother copying from the send to the recieve buffer in bcol_basesmuma_allreduce_intra_fanin_fanout if the pointers are the same. The eliminates a warning about memcpy and aliasing and avoids an unnecessary call to memcpy. Alwasy call CHECK_AND_RELEASE on memsync collectives. There was a call to OBJ_RELEASE on the collective communicator but because CHECK_AND_RECYLCE was never called there was not matching call to OBJ_RELEASE. This caused coll ml to leak communicators. Make allreduce use the sequential collective launcher in coll_ml_inlines.h Just launch the next collective in the component progress. I am a little unsure about this patch. There appears to be some sort of race between collectives that causes buffer exhaustion in some cases (IMB Allreduce is a reproducer). Changing progress to only launch the next bcol seems to resolve the issue but might not be the best fix. Note that I see little-no performance penalty for this change. Fix allreduce when there are extra sources. There was an issue with the buffer offset calculation when there are extra sources. In the case of extra sources == 1 the offset was set to buffer_size (just past the header of the next buffer). I adjusted the buffer size to take into accoun the maximum header size (see the earlier commit that added this) and simplified the offset calculation. Make reduce/allreduce non-blocking. This is required for MPI_Comm_idup to work correctly. This has been tested with various layouts using the ibm testsuite and imb and appears to have the same performance as the old blocking version. Fix allgather for non-contiguous layouts and simplify parsing the topology. Some things in this patch: - There were several comments to the effect that level 0 of the hierarchy MUST contain all of the ranks. At least one function made this assumption but it was not true. I changed the sbgp components and the coll ml initization code to enforce this requirement. - Ensure that hierarchy level 0 has the ranks in the correct scatter gather order. This removes the need for a separate sort list and fixes the offset calculation for allgather. - There were several passes over the hierarchy to determine properties of the hierarchy. I eliminated these extra passes and the memory allocation associated with them and calculate the tree properties on the fly. The same DFS recursion also handles the re-order of level 0. All these changes have been verified with MPI_Allreduce, MPI_Reduce, and MPI_Allgather. All functions now pass all IBM/Open MPI, and IMB tests. coll/ml: correct pointer usage for MPI_BOTTOM Since contiguous datatypes are copied via memcpy (bypassing the convertor) we need to adjust for the lb of the datatype. This corrects problems found testing code that uses MPI_BOTTOM (NULL) as the send pointer. Add fallback collectives for allreduce and reduce. cmr=v1.7.5:reviewer=pasha This commit was SVN r30363.
2014-01-22 19:39:19 +04:00
my_ctl_pointer->flag = ready_flag;
wait_for_peers(my_rank, my_lmsg_ctl_pointer->n_sends, data_buffs,
ready_flag, sequence_number);
} else {
/* takes care of first level recurssive double */
length = parent_lmsg_ctl_pointer->length/
(1<<(parent_lmsg_ctl_pointer->n_sends - 1));
my_lmsg_ctl_pointer->length = length;
my_lmsg_ctl_pointer->offset = parent_lmsg_ctl_pointer->offset;
local_offset = my_lmsg_ctl_pointer->offset;
remote_offset = my_lmsg_ctl_pointer->offset;
while(!IS_SG_DATA_READY(parent_ctl_pointer, ready_flag, sequence_number)) {
opal_progress();
}
mca_bcol_basesmuma_portals_get_msg_fragment_no_eq_h(cs, my_lmsg_ctl_pointer,
parent_lmsg_ctl_pointer,local_offset,
remote_offset, length);
/* signal that I am done reading data from parent */
basesmuma: do not call MB directly opal does not always define MB. It is recommended that opal_atomic_[rw]mb is called instead. We will need to address the cases where these functions are no-ops on weak-memory ordered cpus. cmr=v1.7.5:ticket=trac:4158 This commit was SVN r30534. The following Trac tickets were found above: Ticket 4158 --> https://svn.open-mpi.org/trac/ompi/ticket/4158
2014-02-03 21:01:57 +04:00
opal_atomic_wmb ();
coll/ml: add support for blocking and non-blocking allreduce, reduce, and allgather. The new collectives provide a signifigant performance increase over tuned for small and medium messages. We are initially setting the priority lower than tuned until this has had some time to soak in the trunk. Please set coll_ml_priority to 90 for MTT runs. Credit for this work goes to Manjunath Gorentla Venkata (ORNL), Pavel Shamis (ORNL), and Nathan Hjelm (LANL). Commit details (for reference): Import ORNL's collectives for MPI_Allreduce, MPI_Reduce, and MPI_Allgather. We need to take the basesmuma header into account when calculating the ptpcoll small message thresholds. Add a define to bcol.h indicating the maximum header size so we can take the header into account while not making ptpcoll dependent on information from basesmuma. This resolves an issue with allreduce where ptpcoll overwrites the header of the next buffer in the basesmuma bank. Fix reduce and make a sequential collective launcher in coll_ml_inlines.h The root calculation for reduce was wrong for any root != 0. There are four possibilities for the root: - The root is not the current process but is in the current hierarchy. In this case the root is the index of the global root as specified in the root vector. - The root is not the current process and is not in the next level of the hierarchy. In this case 0 must be the local root since this process will never communicate with the real root. - The root is not the current process but will be in next level of the hierarchy. In this case the current process must be the root. - I am the root. The root is my index. Tested with IMB which rotates the root on every call to MPI_Reduce. Consider IMB the reproducer for the issue this commit solves. Make the bcast algorithm decision an enumerated variable Resolve various asset failures when destructing coll ml requests. Two issues: - Always reset the request to be invalid before returning it to the free list. This will avoid an asset in ompi_request_t's destructor. OMPI_REQUEST_FINI does this (and also releases the fortran handle index). - Never explicitly construct or destruct the superclass of an opal object. This screws up the class function tables and will cause either an assert failure or a segmentation fault when destructing coll ml requests. Cleanup allgather. I removed the duplicate non-blocking and blocking functions and modeled the cleanup after what I found in allreduce. Also cleaned up the code somewhat. Don't bother copying from the send to the recieve buffer in bcol_basesmuma_allreduce_intra_fanin_fanout if the pointers are the same. The eliminates a warning about memcpy and aliasing and avoids an unnecessary call to memcpy. Alwasy call CHECK_AND_RELEASE on memsync collectives. There was a call to OBJ_RELEASE on the collective communicator but because CHECK_AND_RECYLCE was never called there was not matching call to OBJ_RELEASE. This caused coll ml to leak communicators. Make allreduce use the sequential collective launcher in coll_ml_inlines.h Just launch the next collective in the component progress. I am a little unsure about this patch. There appears to be some sort of race between collectives that causes buffer exhaustion in some cases (IMB Allreduce is a reproducer). Changing progress to only launch the next bcol seems to resolve the issue but might not be the best fix. Note that I see little-no performance penalty for this change. Fix allreduce when there are extra sources. There was an issue with the buffer offset calculation when there are extra sources. In the case of extra sources == 1 the offset was set to buffer_size (just past the header of the next buffer). I adjusted the buffer size to take into accoun the maximum header size (see the earlier commit that added this) and simplified the offset calculation. Make reduce/allreduce non-blocking. This is required for MPI_Comm_idup to work correctly. This has been tested with various layouts using the ibm testsuite and imb and appears to have the same performance as the old blocking version. Fix allgather for non-contiguous layouts and simplify parsing the topology. Some things in this patch: - There were several comments to the effect that level 0 of the hierarchy MUST contain all of the ranks. At least one function made this assumption but it was not true. I changed the sbgp components and the coll ml initization code to enforce this requirement. - Ensure that hierarchy level 0 has the ranks in the correct scatter gather order. This removes the need for a separate sort list and fixes the offset calculation for allgather. - There were several passes over the hierarchy to determine properties of the hierarchy. I eliminated these extra passes and the memory allocation associated with them and calculate the tree properties on the fly. The same DFS recursion also handles the re-order of level 0. All these changes have been verified with MPI_Allreduce, MPI_Reduce, and MPI_Allgather. All functions now pass all IBM/Open MPI, and IMB tests. coll/ml: correct pointer usage for MPI_BOTTOM Since contiguous datatypes are copied via memcpy (bypassing the convertor) we need to adjust for the lb of the datatype. This corrects problems found testing code that uses MPI_BOTTOM (NULL) as the send pointer. Add fallback collectives for allreduce and reduce. cmr=v1.7.5:reviewer=pasha This commit was SVN r30363.
2014-01-22 19:39:19 +04:00
my_ctl_pointer->flag = ready_flag;
}
/* time for allgather phase */
input_args->status = ALLGATHER;
BASESMUMA_VERBOSE(5,("Completed %d found it from %d \n",my_rank,src));
while(ready_flag > parent_ctl_pointer->flag);
goto Allgather;
}
}
{
/* this is not who we are looking for,
* mark as false positive so we don't
* poll here again
*/
src_list[src_list_index] = -1;
matched = 0;
goto Probe;
}
Allgather:
BASESMUMA_VERBOSE(5,(" %d Completed Scatter %d times \n", my_rank, completed_scatter));
/* zip it back up - we have already taken care of first level */
global_sg_offset = my_lmsg_ctl_pointer->offset;
/* first level of zip up */
length = 2 * fragment_size/pow_2;
if (!msg_posted) {
rc = PtlEQAlloc(((mca_bcol_basesmuma_portal_proc_info_t*)
cs->portals_info)->ni_h, MAX_PORTAL_EVENTS_IN_Q,
PTL_EQ_HANDLER_NONE, &allgather_eq_h);
/* Posting for all phases of recursive doubling */
extra_src_posts = (my_rank + pow_2 < group_size ) ? 1: 0;
allgather_posts = pow_2_levels - 1;
total_msg_posts = allgather_posts + extra_src_posts ;
mca_bcol_basesmuma_portals_post_msg(cs, my_lmsg_ctl_pointer,
my_userbuf, my_lmsg_ctl_pointer->userbuf_length,
allgather_eq_h, total_msg_posts , blocked_post,
PTL_MD_EVENT_START_DISABLE| PTL_MD_EVENT_END_DISABLE
| PTL_MD_OP_GET | PTL_MD_MANAGE_REMOTE | PTL_MD_TRUNCATE | PTL_MD_EVENT_AUTO_UNLINK_ENABLE
);
msg_posted = true;
}
ready_flag++;
basesmuma: do not call MB directly opal does not always define MB. It is recommended that opal_atomic_[rw]mb is called instead. We will need to address the cases where these functions are no-ops on weak-memory ordered cpus. cmr=v1.7.5:ticket=trac:4158 This commit was SVN r30534. The following Trac tickets were found above: Ticket 4158 --> https://svn.open-mpi.org/trac/ompi/ticket/4158
2014-02-03 21:01:57 +04:00
opal_atomic_wmb ();
coll/ml: add support for blocking and non-blocking allreduce, reduce, and allgather. The new collectives provide a signifigant performance increase over tuned for small and medium messages. We are initially setting the priority lower than tuned until this has had some time to soak in the trunk. Please set coll_ml_priority to 90 for MTT runs. Credit for this work goes to Manjunath Gorentla Venkata (ORNL), Pavel Shamis (ORNL), and Nathan Hjelm (LANL). Commit details (for reference): Import ORNL's collectives for MPI_Allreduce, MPI_Reduce, and MPI_Allgather. We need to take the basesmuma header into account when calculating the ptpcoll small message thresholds. Add a define to bcol.h indicating the maximum header size so we can take the header into account while not making ptpcoll dependent on information from basesmuma. This resolves an issue with allreduce where ptpcoll overwrites the header of the next buffer in the basesmuma bank. Fix reduce and make a sequential collective launcher in coll_ml_inlines.h The root calculation for reduce was wrong for any root != 0. There are four possibilities for the root: - The root is not the current process but is in the current hierarchy. In this case the root is the index of the global root as specified in the root vector. - The root is not the current process and is not in the next level of the hierarchy. In this case 0 must be the local root since this process will never communicate with the real root. - The root is not the current process but will be in next level of the hierarchy. In this case the current process must be the root. - I am the root. The root is my index. Tested with IMB which rotates the root on every call to MPI_Reduce. Consider IMB the reproducer for the issue this commit solves. Make the bcast algorithm decision an enumerated variable Resolve various asset failures when destructing coll ml requests. Two issues: - Always reset the request to be invalid before returning it to the free list. This will avoid an asset in ompi_request_t's destructor. OMPI_REQUEST_FINI does this (and also releases the fortran handle index). - Never explicitly construct or destruct the superclass of an opal object. This screws up the class function tables and will cause either an assert failure or a segmentation fault when destructing coll ml requests. Cleanup allgather. I removed the duplicate non-blocking and blocking functions and modeled the cleanup after what I found in allreduce. Also cleaned up the code somewhat. Don't bother copying from the send to the recieve buffer in bcol_basesmuma_allreduce_intra_fanin_fanout if the pointers are the same. The eliminates a warning about memcpy and aliasing and avoids an unnecessary call to memcpy. Alwasy call CHECK_AND_RELEASE on memsync collectives. There was a call to OBJ_RELEASE on the collective communicator but because CHECK_AND_RECYLCE was never called there was not matching call to OBJ_RELEASE. This caused coll ml to leak communicators. Make allreduce use the sequential collective launcher in coll_ml_inlines.h Just launch the next collective in the component progress. I am a little unsure about this patch. There appears to be some sort of race between collectives that causes buffer exhaustion in some cases (IMB Allreduce is a reproducer). Changing progress to only launch the next bcol seems to resolve the issue but might not be the best fix. Note that I see little-no performance penalty for this change. Fix allreduce when there are extra sources. There was an issue with the buffer offset calculation when there are extra sources. In the case of extra sources == 1 the offset was set to buffer_size (just past the header of the next buffer). I adjusted the buffer size to take into accoun the maximum header size (see the earlier commit that added this) and simplified the offset calculation. Make reduce/allreduce non-blocking. This is required for MPI_Comm_idup to work correctly. This has been tested with various layouts using the ibm testsuite and imb and appears to have the same performance as the old blocking version. Fix allgather for non-contiguous layouts and simplify parsing the topology. Some things in this patch: - There were several comments to the effect that level 0 of the hierarchy MUST contain all of the ranks. At least one function made this assumption but it was not true. I changed the sbgp components and the coll ml initization code to enforce this requirement. - Ensure that hierarchy level 0 has the ranks in the correct scatter gather order. This removes the need for a separate sort list and fixes the offset calculation for allgather. - There were several passes over the hierarchy to determine properties of the hierarchy. I eliminated these extra passes and the memory allocation associated with them and calculate the tree properties on the fly. The same DFS recursion also handles the re-order of level 0. All these changes have been verified with MPI_Allreduce, MPI_Reduce, and MPI_Allgather. All functions now pass all IBM/Open MPI, and IMB tests. coll/ml: correct pointer usage for MPI_BOTTOM Since contiguous datatypes are copied via memcpy (bypassing the convertor) we need to adjust for the lb of the datatype. This corrects problems found testing code that uses MPI_BOTTOM (NULL) as the send pointer. Add fallback collectives for allreduce and reduce. cmr=v1.7.5:reviewer=pasha This commit was SVN r30363.
2014-01-22 19:39:19 +04:00
my_ctl_pointer->flag = ready_flag;
/*
* Recursive doubling allgather implementation
*/
for( i = 1; i < pow_2_levels; i++) {
/* get my partner for this level */
partner = my_rank^(1<<i);
partner_ctl_pointer = data_buffs[partner].ctl_struct;
partner_lmsg_ctl_pointer = (mca_bcol_basesmuma_portal_buf_addr_t*)
data_buffs[partner].payload;
/* Block until partner is at this level of recursive-doubling stage */
while(!IS_SG_DATA_READY(partner_ctl_pointer, ready_flag, sequence_number)) {
opal_progress();
}
assert(partner_ctl_pointer->flag >= ready_flag);
if (partner_lmsg_ctl_pointer->offset < my_lmsg_ctl_pointer->offset) {
global_sg_offset -= length;
local_sg_offset = global_sg_offset;
} else {
local_sg_offset = global_sg_offset + length;
}
BASESMUMA_VERBOSE(10,("Allgather Phase: Get message from process %d, length %d", partner, length));
mca_bcol_basesmuma_portals_get_msg_fragment_no_eq_h(cs, my_lmsg_ctl_pointer,
partner_lmsg_ctl_pointer,local_sg_offset,
local_sg_offset, length);
ready_flag++;
basesmuma: do not call MB directly opal does not always define MB. It is recommended that opal_atomic_[rw]mb is called instead. We will need to address the cases where these functions are no-ops on weak-memory ordered cpus. cmr=v1.7.5:ticket=trac:4158 This commit was SVN r30534. The following Trac tickets were found above: Ticket 4158 --> https://svn.open-mpi.org/trac/ompi/ticket/4158
2014-02-03 21:01:57 +04:00
opal_atomic_wmb ();
coll/ml: add support for blocking and non-blocking allreduce, reduce, and allgather. The new collectives provide a signifigant performance increase over tuned for small and medium messages. We are initially setting the priority lower than tuned until this has had some time to soak in the trunk. Please set coll_ml_priority to 90 for MTT runs. Credit for this work goes to Manjunath Gorentla Venkata (ORNL), Pavel Shamis (ORNL), and Nathan Hjelm (LANL). Commit details (for reference): Import ORNL's collectives for MPI_Allreduce, MPI_Reduce, and MPI_Allgather. We need to take the basesmuma header into account when calculating the ptpcoll small message thresholds. Add a define to bcol.h indicating the maximum header size so we can take the header into account while not making ptpcoll dependent on information from basesmuma. This resolves an issue with allreduce where ptpcoll overwrites the header of the next buffer in the basesmuma bank. Fix reduce and make a sequential collective launcher in coll_ml_inlines.h The root calculation for reduce was wrong for any root != 0. There are four possibilities for the root: - The root is not the current process but is in the current hierarchy. In this case the root is the index of the global root as specified in the root vector. - The root is not the current process and is not in the next level of the hierarchy. In this case 0 must be the local root since this process will never communicate with the real root. - The root is not the current process but will be in next level of the hierarchy. In this case the current process must be the root. - I am the root. The root is my index. Tested with IMB which rotates the root on every call to MPI_Reduce. Consider IMB the reproducer for the issue this commit solves. Make the bcast algorithm decision an enumerated variable Resolve various asset failures when destructing coll ml requests. Two issues: - Always reset the request to be invalid before returning it to the free list. This will avoid an asset in ompi_request_t's destructor. OMPI_REQUEST_FINI does this (and also releases the fortran handle index). - Never explicitly construct or destruct the superclass of an opal object. This screws up the class function tables and will cause either an assert failure or a segmentation fault when destructing coll ml requests. Cleanup allgather. I removed the duplicate non-blocking and blocking functions and modeled the cleanup after what I found in allreduce. Also cleaned up the code somewhat. Don't bother copying from the send to the recieve buffer in bcol_basesmuma_allreduce_intra_fanin_fanout if the pointers are the same. The eliminates a warning about memcpy and aliasing and avoids an unnecessary call to memcpy. Alwasy call CHECK_AND_RELEASE on memsync collectives. There was a call to OBJ_RELEASE on the collective communicator but because CHECK_AND_RECYLCE was never called there was not matching call to OBJ_RELEASE. This caused coll ml to leak communicators. Make allreduce use the sequential collective launcher in coll_ml_inlines.h Just launch the next collective in the component progress. I am a little unsure about this patch. There appears to be some sort of race between collectives that causes buffer exhaustion in some cases (IMB Allreduce is a reproducer). Changing progress to only launch the next bcol seems to resolve the issue but might not be the best fix. Note that I see little-no performance penalty for this change. Fix allreduce when there are extra sources. There was an issue with the buffer offset calculation when there are extra sources. In the case of extra sources == 1 the offset was set to buffer_size (just past the header of the next buffer). I adjusted the buffer size to take into accoun the maximum header size (see the earlier commit that added this) and simplified the offset calculation. Make reduce/allreduce non-blocking. This is required for MPI_Comm_idup to work correctly. This has been tested with various layouts using the ibm testsuite and imb and appears to have the same performance as the old blocking version. Fix allgather for non-contiguous layouts and simplify parsing the topology. Some things in this patch: - There were several comments to the effect that level 0 of the hierarchy MUST contain all of the ranks. At least one function made this assumption but it was not true. I changed the sbgp components and the coll ml initization code to enforce this requirement. - Ensure that hierarchy level 0 has the ranks in the correct scatter gather order. This removes the need for a separate sort list and fixes the offset calculation for allgather. - There were several passes over the hierarchy to determine properties of the hierarchy. I eliminated these extra passes and the memory allocation associated with them and calculate the tree properties on the fly. The same DFS recursion also handles the re-order of level 0. All these changes have been verified with MPI_Allreduce, MPI_Reduce, and MPI_Allgather. All functions now pass all IBM/Open MPI, and IMB tests. coll/ml: correct pointer usage for MPI_BOTTOM Since contiguous datatypes are copied via memcpy (bypassing the convertor) we need to adjust for the lb of the datatype. This corrects problems found testing code that uses MPI_BOTTOM (NULL) as the send pointer. Add fallback collectives for allreduce and reduce. cmr=v1.7.5:reviewer=pasha This commit was SVN r30363.
2014-01-22 19:39:19 +04:00
my_ctl_pointer->flag = ready_flag;
/* Block until partner completed this level of recursive-doubling stage */
while(!IS_SG_DATA_READY(partner_ctl_pointer, ready_flag, sequence_number)) {
opal_progress();
}
/*
* Compute length for next recursive doubling phase
*/
length *= 2;
}
/* If I am source for non-power 2 children wait for them */
/* If I am secondary root then my partner would be real root
* so no need for exchange of data with the extra partner */
extra_partner = my_rank + pow_2 ;
if ((extra_partner < group_size) && (!secondary_root)) {
volatile mca_bcol_basesmuma_header_t *extra_partner_ctl_pointer = NULL;
extra_partner_ctl_pointer = data_buffs[extra_partner].ctl_struct;
/* Block until extra partner has copied data */
while(!IS_SG_DATA_READY(extra_partner_ctl_pointer, ready_flag, sequence_number)) {
opal_progress();
}
}
Release:
/* free the event queue */
rc = PtlEQFree(allgather_eq_h);
if (rc != PTL_OK) {
BASESMUMA_VERBOSE(10,("PtlEQFree() failed: %d )\n",rc));
}
my_ctl_pointer->starting_flag_value++;
input_args->status = FINISHED;
return BCOL_FN_COMPLETE;
}
#endif
/*
* Blocking Portals Scatter Allgather
*
*
*
*
*
*/
int bcol_basesmuma_lmsg_scatter_allgather_portals_bcast(bcol_function_args_t *input_args,
coll_ml_function_t *c_input_args)
{
/* local variables */
int i;
uint64_t length;
int my_rank, parent_rank, src =-1, matched = 0;
int *src_list = NULL;
int group_size = -1, dummy_group_size;
int first_instance=0;
int rc = OMPI_SUCCESS;
int leading_dim, buff_idx, idx;
int count=input_args->count;
size_t pack_len = 0, dt_size =0 ;
volatile int8_t ready_flag;
int flag_offset;
int pow_2, pow_2_levels;
int src_list_index = -1;
uint64_t fragment_size; /* user buffer size */
int sg_matchbits;
/* Input argument variables */
void *my_userbuf = (void*)((unsigned char*)input_args->userbuf);
int64_t sequence_number=input_args->sequence_num;
struct ompi_datatype_t* dtype=input_args->dtype;
/* Extra source variables */
bool secondary_root = false;
int partner = -1, extra_partner = -1;
/* Scatter Allgather offsets */
uint64_t local_sg_offset = 0, global_sg_offset = 0, partner_offset = 0;
/* Portals messaging relevant variables */
mca_bcol_basesmuma_portal_proc_info_t *portals_info;
ptl_handle_eq_t allgather_eq_h;
ptl_event_t allgather_event;
bool blocked_post = false;
bool msg_posted = false;
int total_msg_posts = -1, scatter_posts = -1, allgather_posts = -1, extra_src_posts = -1;
/* OMPI module and component variables */
mca_bcol_basesmuma_component_t *cs = &mca_bcol_basesmuma_component;
mca_bcol_basesmuma_module_t *bcol_module =
(mca_bcol_basesmuma_module_t *) c_input_args->bcol_module;
/* Control structure and payload variables */
volatile mca_bcol_basesmuma_ctl_struct_t **ctl_structs;
volatile mca_bcol_basesmuma_ctl_struct_t *my_ctl_pointer = NULL;
volatile mca_bcol_basesmuma_ctl_struct_t *parent_ctl_pointer = NULL; /* binomial fanout */
volatile mca_bcol_basesmuma_ctl_struct_t *partner_ctl_pointer = NULL; /* recursive double */
/* Make sure there userbuffer is not null */
assert(my_userbuf != NULL);
/* Get portals info*/
portals_info = (mca_bcol_basesmuma_portal_proc_info_t*)cs->portals_info;
/* Get addresing information */
buff_idx = input_args->src_desc->buffer_index;
group_size = bcol_module->colls_no_user_data.size_of_group;
leading_dim=bcol_module->colls_no_user_data.size_of_group;
my_rank = bcol_module->super.sbgp_partner_module->my_index;
idx=SM_ARRAY_INDEX(leading_dim,buff_idx,0);
/* calculate the largest power of two that is smaller than
* or equal to the group size
*/
pow_2_levels = pow_sm_k(2,group_size, &(dummy_group_size));
if( group_size < (1<<pow_2_levels)) {
pow_2_levels--;
}
/* power-of-two group size */
pow_2 = 1<<pow_2_levels;
/* we will work only on packed data - so compute the length*/
ompi_datatype_type_size(dtype, &dt_size);
fragment_size = count*dt_size;
/* grab the ctl buffs */
ctl_structs = (volatile mca_bcol_basesmuma_ctl_struct_t **)
bcol_module->colls_with_user_data.ctl_buffs+idx;
my_ctl_pointer = ctl_structs[my_rank];
if(my_ctl_pointer->sequence_number < sequence_number) {
first_instance = 1;
}
if(first_instance) {
for( i = 0; i < NUM_SIGNAL_FLAGS; i++){
my_ctl_pointer->flags[i] = -1;
}
my_ctl_pointer->index = 1;
my_ctl_pointer->starting_flag_value = 0;
flag_offset = 0;
} else {
my_ctl_pointer->index++;
}
/* increment the starting flag by one and return */
flag_offset = my_ctl_pointer->starting_flag_value;
/*ready_flag = flag_offset + sequence_number + 1;*/
ready_flag = flag_offset + 1;
my_ctl_pointer->sequence_number = sequence_number;
sg_matchbits = sequence_number ;
/* Construct my portal buffer address and copy to payload buffer */
mca_bcol_basesmuma_construct_portal_address(&my_ctl_pointer->portals_buf_addr,
portals_info->portal_id.nid,
portals_info->portal_id.pid,
sg_matchbits,
bcol_module->super.sbgp_partner_module->group_comm->c_contextid);
my_ctl_pointer->portals_buf_addr.userbuf = my_userbuf;
my_ctl_pointer->portals_buf_addr.userbuf_length = fragment_size;
if (input_args->root_flag) {
my_ctl_pointer->offset = 0;
my_ctl_pointer->n_sends = pow_2_levels;
my_ctl_pointer->length = fragment_size;
rc = PtlEQAlloc(((mca_bcol_basesmuma_portal_proc_info_t*)
cs->portals_info)->ni_h, MAX_PORTAL_EVENTS_IN_Q,
PTL_EQ_HANDLER_NONE, &allgather_eq_h);
if (rc != PTL_OK) {
BASESMUMA_VERBOSE(10,( "PtlEQAlloc() failed: %d \n",rc));
goto Release;
}
/* Compute number of posts required */
if (my_rank >= pow_2) {
/* I am root and my rank is greater than pow_2, I will hand
* over to rank (that is < pow_2) to act as secondary root
*/
total_msg_posts = 1;
}
else {
extra_src_posts = (my_rank + pow_2 < group_size ) ? 1: 0;
scatter_posts = my_ctl_pointer->n_sends;
allgather_posts = pow_2_levels - 1;
total_msg_posts = scatter_posts + allgather_posts + extra_src_posts ;
}
mca_bcol_basesmuma_portals_post_msg(cs,
&my_ctl_pointer->portals_buf_addr,
my_userbuf, fragment_size, allgather_eq_h,
total_msg_posts,
blocked_post,
PTL_MD_EVENT_START_DISABLE| PTL_MD_EVENT_END_DISABLE | PTL_MD_OP_GET | PTL_MD_MANAGE_REMOTE | PTL_MD_TRUNCATE | PTL_MD_EVENT_AUTO_UNLINK_ENABLE);
msg_posted = true ;
/* important that these be set before my children
* see the ready flag raised
*/
basesmuma: do not call MB directly opal does not always define MB. It is recommended that opal_atomic_[rw]mb is called instead. We will need to address the cases where these functions are no-ops on weak-memory ordered cpus. cmr=v1.7.5:ticket=trac:4158 This commit was SVN r30534. The following Trac tickets were found above: Ticket 4158 --> https://svn.open-mpi.org/trac/ompi/ticket/4158
2014-02-03 21:01:57 +04:00
opal_atomic_wmb ();
coll/ml: add support for blocking and non-blocking allreduce, reduce, and allgather. The new collectives provide a signifigant performance increase over tuned for small and medium messages. We are initially setting the priority lower than tuned until this has had some time to soak in the trunk. Please set coll_ml_priority to 90 for MTT runs. Credit for this work goes to Manjunath Gorentla Venkata (ORNL), Pavel Shamis (ORNL), and Nathan Hjelm (LANL). Commit details (for reference): Import ORNL's collectives for MPI_Allreduce, MPI_Reduce, and MPI_Allgather. We need to take the basesmuma header into account when calculating the ptpcoll small message thresholds. Add a define to bcol.h indicating the maximum header size so we can take the header into account while not making ptpcoll dependent on information from basesmuma. This resolves an issue with allreduce where ptpcoll overwrites the header of the next buffer in the basesmuma bank. Fix reduce and make a sequential collective launcher in coll_ml_inlines.h The root calculation for reduce was wrong for any root != 0. There are four possibilities for the root: - The root is not the current process but is in the current hierarchy. In this case the root is the index of the global root as specified in the root vector. - The root is not the current process and is not in the next level of the hierarchy. In this case 0 must be the local root since this process will never communicate with the real root. - The root is not the current process but will be in next level of the hierarchy. In this case the current process must be the root. - I am the root. The root is my index. Tested with IMB which rotates the root on every call to MPI_Reduce. Consider IMB the reproducer for the issue this commit solves. Make the bcast algorithm decision an enumerated variable Resolve various asset failures when destructing coll ml requests. Two issues: - Always reset the request to be invalid before returning it to the free list. This will avoid an asset in ompi_request_t's destructor. OMPI_REQUEST_FINI does this (and also releases the fortran handle index). - Never explicitly construct or destruct the superclass of an opal object. This screws up the class function tables and will cause either an assert failure or a segmentation fault when destructing coll ml requests. Cleanup allgather. I removed the duplicate non-blocking and blocking functions and modeled the cleanup after what I found in allreduce. Also cleaned up the code somewhat. Don't bother copying from the send to the recieve buffer in bcol_basesmuma_allreduce_intra_fanin_fanout if the pointers are the same. The eliminates a warning about memcpy and aliasing and avoids an unnecessary call to memcpy. Alwasy call CHECK_AND_RELEASE on memsync collectives. There was a call to OBJ_RELEASE on the collective communicator but because CHECK_AND_RECYLCE was never called there was not matching call to OBJ_RELEASE. This caused coll ml to leak communicators. Make allreduce use the sequential collective launcher in coll_ml_inlines.h Just launch the next collective in the component progress. I am a little unsure about this patch. There appears to be some sort of race between collectives that causes buffer exhaustion in some cases (IMB Allreduce is a reproducer). Changing progress to only launch the next bcol seems to resolve the issue but might not be the best fix. Note that I see little-no performance penalty for this change. Fix allreduce when there are extra sources. There was an issue with the buffer offset calculation when there are extra sources. In the case of extra sources == 1 the offset was set to buffer_size (just past the header of the next buffer). I adjusted the buffer size to take into accoun the maximum header size (see the earlier commit that added this) and simplified the offset calculation. Make reduce/allreduce non-blocking. This is required for MPI_Comm_idup to work correctly. This has been tested with various layouts using the ibm testsuite and imb and appears to have the same performance as the old blocking version. Fix allgather for non-contiguous layouts and simplify parsing the topology. Some things in this patch: - There were several comments to the effect that level 0 of the hierarchy MUST contain all of the ranks. At least one function made this assumption but it was not true. I changed the sbgp components and the coll ml initization code to enforce this requirement. - Ensure that hierarchy level 0 has the ranks in the correct scatter gather order. This removes the need for a separate sort list and fixes the offset calculation for allgather. - There were several passes over the hierarchy to determine properties of the hierarchy. I eliminated these extra passes and the memory allocation associated with them and calculate the tree properties on the fly. The same DFS recursion also handles the re-order of level 0. All these changes have been verified with MPI_Allreduce, MPI_Reduce, and MPI_Allgather. All functions now pass all IBM/Open MPI, and IMB tests. coll/ml: correct pointer usage for MPI_BOTTOM Since contiguous datatypes are copied via memcpy (bypassing the convertor) we need to adjust for the lb of the datatype. This corrects problems found testing code that uses MPI_BOTTOM (NULL) as the send pointer. Add fallback collectives for allreduce and reduce. cmr=v1.7.5:reviewer=pasha This commit was SVN r30363.
2014-01-22 19:39:19 +04:00
my_ctl_pointer->flags[BCAST_FLAGS] = ready_flag;
BASESMUMA_VERBOSE(1,("I am the root(ctl_pointer %x) of the data flag value %d",my_ctl_pointer, my_ctl_pointer->flag));
/* Wait for my scatter partner */
if (my_rank >= pow_2) {
int scatter_partner = -1;
volatile mca_bcol_basesmuma_ctl_struct_t *scatter_partner_ctl_pointer = NULL;
scatter_partner = my_rank - pow_2;
scatter_partner_ctl_pointer =
ctl_structs[scatter_partner];
while(!IS_SG_DATA_READY(scatter_partner_ctl_pointer, ready_flag,
sequence_number)){
SCATTER_WAIT_FOR_EXTRA:
opal_progress();
}
goto Release;
}
else {
wait_for_peers_nb(my_rank, my_ctl_pointer->n_sends, ctl_structs,
ready_flag, sequence_number);
}
goto Allgather;
}
Extra :
if( my_rank >= pow_2 ) {
parent_rank = my_rank & (pow_2-1);
parent_ctl_pointer = ctl_structs[parent_rank];
ready_flag = ready_flag + pow_2_levels;
while(!IS_SG_DATA_READY(parent_ctl_pointer, ready_flag, sequence_number)) {
opal_progress();
}
mca_bcol_basesmuma_portals_get_msg_fragment_no_eq_h(cs, &my_ctl_pointer->portals_buf_addr,
&parent_ctl_pointer->portals_buf_addr, 0,
0, fragment_size);
my_ctl_pointer->flags[BCAST_FLAG] = ready_flag;
goto Release;
}
Scatter:
/* compute the list of possible sources */
src_list = (int *) malloc(sizeof(int) * (pow_2_levels + 1));
for( i = 0; i < pow_2_levels; i++) {
src_list[i] = my_rank ^ (1<<i);
}
/* My source might be process > pow_2 */
if ((my_rank + pow_2) < group_size) {
src_list[i] = my_rank + pow_2;
} else {
src_list[i] = -1;
}
Probe:
/* If I am not the root, then poll on possible "senders'" control structs */
/* For portals we block for now */
while (!matched) {
/* Shared memory iprobe */
SG_LARGE_MSG_NB_PROBE(src_list, pow_2_levels + 1,
src_list_index, matched, src, ctl_structs,
parent_ctl_pointer, ready_flag, sequence_number);
}
BASESMUMA_VERBOSE(1,("Scatter : Im non-root match received"));
/* If I am a secondary root */
if ((matched) && (src == pow_2 + my_rank)) {
volatile mca_bcol_basesmuma_ctl_struct_t *extra_src_ctl_pointer = NULL;
secondary_root = true;
BASESMUMA_VERBOSE(10,("I am the secondary root for the data"));
my_ctl_pointer->offset = 0;
my_ctl_pointer->n_sends = pow_2_levels;
my_ctl_pointer->length = fragment_size;
extra_src_ctl_pointer = ctl_structs[src];
/* create an event queue for the incoming buffer */
rc = PtlEQAlloc(((mca_bcol_basesmuma_portal_proc_info_t*)
cs->portals_info)->ni_h, MAX_PORTAL_EVENTS_IN_Q,
PTL_EQ_HANDLER_NONE, &allgather_eq_h);
if (rc != PTL_OK) {
BASESMUMA_VERBOSE(10,( "PtlEQAlloc() failed: %d \n",rc));
goto Release;
}
mca_bcol_basesmuma_portals_get_msg_fragment_no_eq_h(cs,
&my_ctl_pointer->portals_buf_addr,
&extra_src_ctl_pointer->portals_buf_addr, 0,
0, fragment_size);
extra_src_posts = 0;
scatter_posts = my_ctl_pointer->n_sends;
allgather_posts = pow_2_levels - 1;
total_msg_posts = scatter_posts + allgather_posts + extra_src_posts ;
mca_bcol_basesmuma_portals_post_msg(cs,
&my_ctl_pointer->portals_buf_addr,
my_userbuf, fragment_size, allgather_eq_h,
total_msg_posts,
blocked_post,
PTL_MD_EVENT_START_DISABLE| PTL_MD_EVENT_END_DISABLE | PTL_MD_OP_GET
| PTL_MD_MANAGE_REMOTE | PTL_MD_TRUNCATE | PTL_MD_EVENT_AUTO_UNLINK_ENABLE);
msg_posted = true ;
/* important that these be set before my children
* see the ready flag raised
*/
basesmuma: do not call MB directly opal does not always define MB. It is recommended that opal_atomic_[rw]mb is called instead. We will need to address the cases where these functions are no-ops on weak-memory ordered cpus. cmr=v1.7.5:ticket=trac:4158 This commit was SVN r30534. The following Trac tickets were found above: Ticket 4158 --> https://svn.open-mpi.org/trac/ompi/ticket/4158
2014-02-03 21:01:57 +04:00
opal_atomic_wmb ();
coll/ml: add support for blocking and non-blocking allreduce, reduce, and allgather. The new collectives provide a signifigant performance increase over tuned for small and medium messages. We are initially setting the priority lower than tuned until this has had some time to soak in the trunk. Please set coll_ml_priority to 90 for MTT runs. Credit for this work goes to Manjunath Gorentla Venkata (ORNL), Pavel Shamis (ORNL), and Nathan Hjelm (LANL). Commit details (for reference): Import ORNL's collectives for MPI_Allreduce, MPI_Reduce, and MPI_Allgather. We need to take the basesmuma header into account when calculating the ptpcoll small message thresholds. Add a define to bcol.h indicating the maximum header size so we can take the header into account while not making ptpcoll dependent on information from basesmuma. This resolves an issue with allreduce where ptpcoll overwrites the header of the next buffer in the basesmuma bank. Fix reduce and make a sequential collective launcher in coll_ml_inlines.h The root calculation for reduce was wrong for any root != 0. There are four possibilities for the root: - The root is not the current process but is in the current hierarchy. In this case the root is the index of the global root as specified in the root vector. - The root is not the current process and is not in the next level of the hierarchy. In this case 0 must be the local root since this process will never communicate with the real root. - The root is not the current process but will be in next level of the hierarchy. In this case the current process must be the root. - I am the root. The root is my index. Tested with IMB which rotates the root on every call to MPI_Reduce. Consider IMB the reproducer for the issue this commit solves. Make the bcast algorithm decision an enumerated variable Resolve various asset failures when destructing coll ml requests. Two issues: - Always reset the request to be invalid before returning it to the free list. This will avoid an asset in ompi_request_t's destructor. OMPI_REQUEST_FINI does this (and also releases the fortran handle index). - Never explicitly construct or destruct the superclass of an opal object. This screws up the class function tables and will cause either an assert failure or a segmentation fault when destructing coll ml requests. Cleanup allgather. I removed the duplicate non-blocking and blocking functions and modeled the cleanup after what I found in allreduce. Also cleaned up the code somewhat. Don't bother copying from the send to the recieve buffer in bcol_basesmuma_allreduce_intra_fanin_fanout if the pointers are the same. The eliminates a warning about memcpy and aliasing and avoids an unnecessary call to memcpy. Alwasy call CHECK_AND_RELEASE on memsync collectives. There was a call to OBJ_RELEASE on the collective communicator but because CHECK_AND_RECYLCE was never called there was not matching call to OBJ_RELEASE. This caused coll ml to leak communicators. Make allreduce use the sequential collective launcher in coll_ml_inlines.h Just launch the next collective in the component progress. I am a little unsure about this patch. There appears to be some sort of race between collectives that causes buffer exhaustion in some cases (IMB Allreduce is a reproducer). Changing progress to only launch the next bcol seems to resolve the issue but might not be the best fix. Note that I see little-no performance penalty for this change. Fix allreduce when there are extra sources. There was an issue with the buffer offset calculation when there are extra sources. In the case of extra sources == 1 the offset was set to buffer_size (just past the header of the next buffer). I adjusted the buffer size to take into accoun the maximum header size (see the earlier commit that added this) and simplified the offset calculation. Make reduce/allreduce non-blocking. This is required for MPI_Comm_idup to work correctly. This has been tested with various layouts using the ibm testsuite and imb and appears to have the same performance as the old blocking version. Fix allgather for non-contiguous layouts and simplify parsing the topology. Some things in this patch: - There were several comments to the effect that level 0 of the hierarchy MUST contain all of the ranks. At least one function made this assumption but it was not true. I changed the sbgp components and the coll ml initization code to enforce this requirement. - Ensure that hierarchy level 0 has the ranks in the correct scatter gather order. This removes the need for a separate sort list and fixes the offset calculation for allgather. - There were several passes over the hierarchy to determine properties of the hierarchy. I eliminated these extra passes and the memory allocation associated with them and calculate the tree properties on the fly. The same DFS recursion also handles the re-order of level 0. All these changes have been verified with MPI_Allreduce, MPI_Reduce, and MPI_Allgather. All functions now pass all IBM/Open MPI, and IMB tests. coll/ml: correct pointer usage for MPI_BOTTOM Since contiguous datatypes are copied via memcpy (bypassing the convertor) we need to adjust for the lb of the datatype. This corrects problems found testing code that uses MPI_BOTTOM (NULL) as the send pointer. Add fallback collectives for allreduce and reduce. cmr=v1.7.5:reviewer=pasha This commit was SVN r30363.
2014-01-22 19:39:19 +04:00
my_ctl_pointer->flags[BCAST_FLAG] = ready_flag;
wait_for_peers_nb(my_rank, my_ctl_pointer->n_sends, ctl_structs,
ready_flag, sequence_number);
goto Allgather;
}
/* we need to see whether this is really
* who we are looking for
*/
for( i = 0; i < parent_ctl_pointer->n_sends; i++) {
uint64_t local_offset = 0;
uint64_t remote_offset = 0;
BASESMUMA_VERBOSE(1,("%d found it from %d \n",my_rank,src));
if( my_rank == (src^(1<<i))) {
parent_ctl_pointer = ctl_structs[src];
/* we found our root within the group ... */
BASESMUMA_VERBOSE(10,("Shared memory probe was matched, the root is %d", src));
my_ctl_pointer->n_sends = i;
/* Am I source for other process during scatter phase */
if ( i > 0) {
/* compute the size of the chunk to copy */
length = (parent_ctl_pointer->length)/
(1<<(parent_ctl_pointer->n_sends - my_ctl_pointer->n_sends));
my_ctl_pointer->length = length;
my_ctl_pointer->offset =
parent_ctl_pointer->offset + length;
local_offset = my_ctl_pointer->offset;
remote_offset = parent_ctl_pointer->offset + length;
mca_bcol_basesmuma_portals_get_msg_fragment_no_eq_h(cs, &my_ctl_pointer->portals_buf_addr,
&parent_ctl_pointer->portals_buf_addr,local_offset,
remote_offset, length);
rc = PtlEQAlloc(((mca_bcol_basesmuma_portal_proc_info_t*)
cs->portals_info)->ni_h, MAX_PORTAL_EVENTS_IN_Q,
PTL_EQ_HANDLER_NONE,
&allgather_eq_h);
if (rc != PTL_OK) {
BASESMUMA_VERBOSE(10,( "PtlEQAlloc() failed: %d \n",rc));
goto Release;
}
/* Now post the message for other children to read */
extra_src_posts = (my_rank + pow_2 < group_size ) ? 1: 0;
scatter_posts = my_ctl_pointer->n_sends;
allgather_posts = pow_2_levels - 1;
total_msg_posts = scatter_posts + allgather_posts + extra_src_posts ;
mca_bcol_basesmuma_portals_post_msg(cs, &my_ctl_pointer->portals_buf_addr,
my_userbuf, my_ctl_pointer->portals_buf_addr.userbuf_length,
allgather_eq_h,
total_msg_posts,
blocked_post,
PTL_MD_EVENT_START_DISABLE| PTL_MD_EVENT_END_DISABLE
| PTL_MD_OP_GET | PTL_MD_MANAGE_REMOTE | PTL_MD_TRUNCATE | PTL_MD_EVENT_AUTO_UNLINK_ENABLE
);
msg_posted = true;
/* set the memory barrier to ensure completion */
basesmuma: do not call MB directly opal does not always define MB. It is recommended that opal_atomic_[rw]mb is called instead. We will need to address the cases where these functions are no-ops on weak-memory ordered cpus. cmr=v1.7.5:ticket=trac:4158 This commit was SVN r30534. The following Trac tickets were found above: Ticket 4158 --> https://svn.open-mpi.org/trac/ompi/ticket/4158
2014-02-03 21:01:57 +04:00
opal_atomic_wmb ();
coll/ml: add support for blocking and non-blocking allreduce, reduce, and allgather. The new collectives provide a signifigant performance increase over tuned for small and medium messages. We are initially setting the priority lower than tuned until this has had some time to soak in the trunk. Please set coll_ml_priority to 90 for MTT runs. Credit for this work goes to Manjunath Gorentla Venkata (ORNL), Pavel Shamis (ORNL), and Nathan Hjelm (LANL). Commit details (for reference): Import ORNL's collectives for MPI_Allreduce, MPI_Reduce, and MPI_Allgather. We need to take the basesmuma header into account when calculating the ptpcoll small message thresholds. Add a define to bcol.h indicating the maximum header size so we can take the header into account while not making ptpcoll dependent on information from basesmuma. This resolves an issue with allreduce where ptpcoll overwrites the header of the next buffer in the basesmuma bank. Fix reduce and make a sequential collective launcher in coll_ml_inlines.h The root calculation for reduce was wrong for any root != 0. There are four possibilities for the root: - The root is not the current process but is in the current hierarchy. In this case the root is the index of the global root as specified in the root vector. - The root is not the current process and is not in the next level of the hierarchy. In this case 0 must be the local root since this process will never communicate with the real root. - The root is not the current process but will be in next level of the hierarchy. In this case the current process must be the root. - I am the root. The root is my index. Tested with IMB which rotates the root on every call to MPI_Reduce. Consider IMB the reproducer for the issue this commit solves. Make the bcast algorithm decision an enumerated variable Resolve various asset failures when destructing coll ml requests. Two issues: - Always reset the request to be invalid before returning it to the free list. This will avoid an asset in ompi_request_t's destructor. OMPI_REQUEST_FINI does this (and also releases the fortran handle index). - Never explicitly construct or destruct the superclass of an opal object. This screws up the class function tables and will cause either an assert failure or a segmentation fault when destructing coll ml requests. Cleanup allgather. I removed the duplicate non-blocking and blocking functions and modeled the cleanup after what I found in allreduce. Also cleaned up the code somewhat. Don't bother copying from the send to the recieve buffer in bcol_basesmuma_allreduce_intra_fanin_fanout if the pointers are the same. The eliminates a warning about memcpy and aliasing and avoids an unnecessary call to memcpy. Alwasy call CHECK_AND_RELEASE on memsync collectives. There was a call to OBJ_RELEASE on the collective communicator but because CHECK_AND_RECYLCE was never called there was not matching call to OBJ_RELEASE. This caused coll ml to leak communicators. Make allreduce use the sequential collective launcher in coll_ml_inlines.h Just launch the next collective in the component progress. I am a little unsure about this patch. There appears to be some sort of race between collectives that causes buffer exhaustion in some cases (IMB Allreduce is a reproducer). Changing progress to only launch the next bcol seems to resolve the issue but might not be the best fix. Note that I see little-no performance penalty for this change. Fix allreduce when there are extra sources. There was an issue with the buffer offset calculation when there are extra sources. In the case of extra sources == 1 the offset was set to buffer_size (just past the header of the next buffer). I adjusted the buffer size to take into accoun the maximum header size (see the earlier commit that added this) and simplified the offset calculation. Make reduce/allreduce non-blocking. This is required for MPI_Comm_idup to work correctly. This has been tested with various layouts using the ibm testsuite and imb and appears to have the same performance as the old blocking version. Fix allgather for non-contiguous layouts and simplify parsing the topology. Some things in this patch: - There were several comments to the effect that level 0 of the hierarchy MUST contain all of the ranks. At least one function made this assumption but it was not true. I changed the sbgp components and the coll ml initization code to enforce this requirement. - Ensure that hierarchy level 0 has the ranks in the correct scatter gather order. This removes the need for a separate sort list and fixes the offset calculation for allgather. - There were several passes over the hierarchy to determine properties of the hierarchy. I eliminated these extra passes and the memory allocation associated with them and calculate the tree properties on the fly. The same DFS recursion also handles the re-order of level 0. All these changes have been verified with MPI_Allreduce, MPI_Reduce, and MPI_Allgather. All functions now pass all IBM/Open MPI, and IMB tests. coll/ml: correct pointer usage for MPI_BOTTOM Since contiguous datatypes are copied via memcpy (bypassing the convertor) we need to adjust for the lb of the datatype. This corrects problems found testing code that uses MPI_BOTTOM (NULL) as the send pointer. Add fallback collectives for allreduce and reduce. cmr=v1.7.5:reviewer=pasha This commit was SVN r30363.
2014-01-22 19:39:19 +04:00
/* signal that I am done */
my_ctl_pointer->flags[BCAST_FLAG] = ready_flag;
wait_for_peers_nb(my_rank, my_ctl_pointer->n_sends, ctl_structs,
ready_flag, sequence_number);
} else {
/* takes care of first level recurssive double */
length = parent_ctl_pointer->length/
(1<<(parent_ctl_pointer->n_sends - 1));
my_ctl_pointer->length = length;
my_ctl_pointer->offset = parent_ctl_pointer->offset;
local_offset = my_ctl_pointer->offset;
remote_offset = my_ctl_pointer->offset;
while(!IS_SG_DATA_READY(parent_ctl_pointer, ready_flag, sequence_number)) {
opal_progress();
}
mca_bcol_basesmuma_portals_get_msg_fragment_no_eq_h(cs,
&my_ctl_pointer->portals_buf_addr,
&parent_ctl_pointer->portals_buf_addr, local_offset,
remote_offset, length);
/* signal that I am done reading data from parent */
basesmuma: do not call MB directly opal does not always define MB. It is recommended that opal_atomic_[rw]mb is called instead. We will need to address the cases where these functions are no-ops on weak-memory ordered cpus. cmr=v1.7.5:ticket=trac:4158 This commit was SVN r30534. The following Trac tickets were found above: Ticket 4158 --> https://svn.open-mpi.org/trac/ompi/ticket/4158
2014-02-03 21:01:57 +04:00
opal_atomic_wmb ();
coll/ml: add support for blocking and non-blocking allreduce, reduce, and allgather. The new collectives provide a signifigant performance increase over tuned for small and medium messages. We are initially setting the priority lower than tuned until this has had some time to soak in the trunk. Please set coll_ml_priority to 90 for MTT runs. Credit for this work goes to Manjunath Gorentla Venkata (ORNL), Pavel Shamis (ORNL), and Nathan Hjelm (LANL). Commit details (for reference): Import ORNL's collectives for MPI_Allreduce, MPI_Reduce, and MPI_Allgather. We need to take the basesmuma header into account when calculating the ptpcoll small message thresholds. Add a define to bcol.h indicating the maximum header size so we can take the header into account while not making ptpcoll dependent on information from basesmuma. This resolves an issue with allreduce where ptpcoll overwrites the header of the next buffer in the basesmuma bank. Fix reduce and make a sequential collective launcher in coll_ml_inlines.h The root calculation for reduce was wrong for any root != 0. There are four possibilities for the root: - The root is not the current process but is in the current hierarchy. In this case the root is the index of the global root as specified in the root vector. - The root is not the current process and is not in the next level of the hierarchy. In this case 0 must be the local root since this process will never communicate with the real root. - The root is not the current process but will be in next level of the hierarchy. In this case the current process must be the root. - I am the root. The root is my index. Tested with IMB which rotates the root on every call to MPI_Reduce. Consider IMB the reproducer for the issue this commit solves. Make the bcast algorithm decision an enumerated variable Resolve various asset failures when destructing coll ml requests. Two issues: - Always reset the request to be invalid before returning it to the free list. This will avoid an asset in ompi_request_t's destructor. OMPI_REQUEST_FINI does this (and also releases the fortran handle index). - Never explicitly construct or destruct the superclass of an opal object. This screws up the class function tables and will cause either an assert failure or a segmentation fault when destructing coll ml requests. Cleanup allgather. I removed the duplicate non-blocking and blocking functions and modeled the cleanup after what I found in allreduce. Also cleaned up the code somewhat. Don't bother copying from the send to the recieve buffer in bcol_basesmuma_allreduce_intra_fanin_fanout if the pointers are the same. The eliminates a warning about memcpy and aliasing and avoids an unnecessary call to memcpy. Alwasy call CHECK_AND_RELEASE on memsync collectives. There was a call to OBJ_RELEASE on the collective communicator but because CHECK_AND_RECYLCE was never called there was not matching call to OBJ_RELEASE. This caused coll ml to leak communicators. Make allreduce use the sequential collective launcher in coll_ml_inlines.h Just launch the next collective in the component progress. I am a little unsure about this patch. There appears to be some sort of race between collectives that causes buffer exhaustion in some cases (IMB Allreduce is a reproducer). Changing progress to only launch the next bcol seems to resolve the issue but might not be the best fix. Note that I see little-no performance penalty for this change. Fix allreduce when there are extra sources. There was an issue with the buffer offset calculation when there are extra sources. In the case of extra sources == 1 the offset was set to buffer_size (just past the header of the next buffer). I adjusted the buffer size to take into accoun the maximum header size (see the earlier commit that added this) and simplified the offset calculation. Make reduce/allreduce non-blocking. This is required for MPI_Comm_idup to work correctly. This has been tested with various layouts using the ibm testsuite and imb and appears to have the same performance as the old blocking version. Fix allgather for non-contiguous layouts and simplify parsing the topology. Some things in this patch: - There were several comments to the effect that level 0 of the hierarchy MUST contain all of the ranks. At least one function made this assumption but it was not true. I changed the sbgp components and the coll ml initization code to enforce this requirement. - Ensure that hierarchy level 0 has the ranks in the correct scatter gather order. This removes the need for a separate sort list and fixes the offset calculation for allgather. - There were several passes over the hierarchy to determine properties of the hierarchy. I eliminated these extra passes and the memory allocation associated with them and calculate the tree properties on the fly. The same DFS recursion also handles the re-order of level 0. All these changes have been verified with MPI_Allreduce, MPI_Reduce, and MPI_Allgather. All functions now pass all IBM/Open MPI, and IMB tests. coll/ml: correct pointer usage for MPI_BOTTOM Since contiguous datatypes are copied via memcpy (bypassing the convertor) we need to adjust for the lb of the datatype. This corrects problems found testing code that uses MPI_BOTTOM (NULL) as the send pointer. Add fallback collectives for allreduce and reduce. cmr=v1.7.5:reviewer=pasha This commit was SVN r30363.
2014-01-22 19:39:19 +04:00
my_ctl_pointer->flags[BCAST_FLAG] = ready_flag;
}
/* time for allgather phase */
input_args->status = ALLGATHER;
BASESMUMA_VERBOSE(5,("Completed %d found it from %d \n",my_rank,src));
while(ready_flag > parent_ctl_pointer->flags[BCAST_FLAG]);
goto Allgather;
}
}
{
/* this is not who we are looking for,
* mark as false positive so we don't
* poll here again
*/
src_list[src_list_index] = -1;
matched = 0;
goto Probe;
}
Allgather:
BASESMUMA_VERBOSE(5,(" %d Completed Scatter %d times \n", my_rank, completed_scatter));
/* zip it back up - we have already taken care of first level */
global_sg_offset = my_ctl_pointer->offset;
/* first level of zip up */
length = 2 * fragment_size/pow_2;
if (!msg_posted) {
rc = PtlEQAlloc(((mca_bcol_basesmuma_portal_proc_info_t*)
cs->portals_info)->ni_h, MAX_PORTAL_EVENTS_IN_Q,
PTL_EQ_HANDLER_NONE, &allgather_eq_h);
/* Posting for all phases of recursive doubling */
extra_src_posts = (my_rank + pow_2 < group_size ) ? 1: 0;
allgather_posts = pow_2_levels - 1;
total_msg_posts = allgather_posts + extra_src_posts ;
mca_bcol_basesmuma_portals_post_msg(cs, &my_ctl_pointer->portals_buf_addr,
my_userbuf, my_ctl_pointer->portals_buf_addr.userbuf_length,
allgather_eq_h, total_msg_posts , blocked_post,
PTL_MD_EVENT_START_DISABLE| PTL_MD_EVENT_END_DISABLE
| PTL_MD_OP_GET | PTL_MD_MANAGE_REMOTE | PTL_MD_TRUNCATE | PTL_MD_EVENT_AUTO_UNLINK_ENABLE
);
msg_posted = true;
}
ready_flag++;
basesmuma: do not call MB directly opal does not always define MB. It is recommended that opal_atomic_[rw]mb is called instead. We will need to address the cases where these functions are no-ops on weak-memory ordered cpus. cmr=v1.7.5:ticket=trac:4158 This commit was SVN r30534. The following Trac tickets were found above: Ticket 4158 --> https://svn.open-mpi.org/trac/ompi/ticket/4158
2014-02-03 21:01:57 +04:00
opal_atomic_wmb ();
coll/ml: add support for blocking and non-blocking allreduce, reduce, and allgather. The new collectives provide a signifigant performance increase over tuned for small and medium messages. We are initially setting the priority lower than tuned until this has had some time to soak in the trunk. Please set coll_ml_priority to 90 for MTT runs. Credit for this work goes to Manjunath Gorentla Venkata (ORNL), Pavel Shamis (ORNL), and Nathan Hjelm (LANL). Commit details (for reference): Import ORNL's collectives for MPI_Allreduce, MPI_Reduce, and MPI_Allgather. We need to take the basesmuma header into account when calculating the ptpcoll small message thresholds. Add a define to bcol.h indicating the maximum header size so we can take the header into account while not making ptpcoll dependent on information from basesmuma. This resolves an issue with allreduce where ptpcoll overwrites the header of the next buffer in the basesmuma bank. Fix reduce and make a sequential collective launcher in coll_ml_inlines.h The root calculation for reduce was wrong for any root != 0. There are four possibilities for the root: - The root is not the current process but is in the current hierarchy. In this case the root is the index of the global root as specified in the root vector. - The root is not the current process and is not in the next level of the hierarchy. In this case 0 must be the local root since this process will never communicate with the real root. - The root is not the current process but will be in next level of the hierarchy. In this case the current process must be the root. - I am the root. The root is my index. Tested with IMB which rotates the root on every call to MPI_Reduce. Consider IMB the reproducer for the issue this commit solves. Make the bcast algorithm decision an enumerated variable Resolve various asset failures when destructing coll ml requests. Two issues: - Always reset the request to be invalid before returning it to the free list. This will avoid an asset in ompi_request_t's destructor. OMPI_REQUEST_FINI does this (and also releases the fortran handle index). - Never explicitly construct or destruct the superclass of an opal object. This screws up the class function tables and will cause either an assert failure or a segmentation fault when destructing coll ml requests. Cleanup allgather. I removed the duplicate non-blocking and blocking functions and modeled the cleanup after what I found in allreduce. Also cleaned up the code somewhat. Don't bother copying from the send to the recieve buffer in bcol_basesmuma_allreduce_intra_fanin_fanout if the pointers are the same. The eliminates a warning about memcpy and aliasing and avoids an unnecessary call to memcpy. Alwasy call CHECK_AND_RELEASE on memsync collectives. There was a call to OBJ_RELEASE on the collective communicator but because CHECK_AND_RECYLCE was never called there was not matching call to OBJ_RELEASE. This caused coll ml to leak communicators. Make allreduce use the sequential collective launcher in coll_ml_inlines.h Just launch the next collective in the component progress. I am a little unsure about this patch. There appears to be some sort of race between collectives that causes buffer exhaustion in some cases (IMB Allreduce is a reproducer). Changing progress to only launch the next bcol seems to resolve the issue but might not be the best fix. Note that I see little-no performance penalty for this change. Fix allreduce when there are extra sources. There was an issue with the buffer offset calculation when there are extra sources. In the case of extra sources == 1 the offset was set to buffer_size (just past the header of the next buffer). I adjusted the buffer size to take into accoun the maximum header size (see the earlier commit that added this) and simplified the offset calculation. Make reduce/allreduce non-blocking. This is required for MPI_Comm_idup to work correctly. This has been tested with various layouts using the ibm testsuite and imb and appears to have the same performance as the old blocking version. Fix allgather for non-contiguous layouts and simplify parsing the topology. Some things in this patch: - There were several comments to the effect that level 0 of the hierarchy MUST contain all of the ranks. At least one function made this assumption but it was not true. I changed the sbgp components and the coll ml initization code to enforce this requirement. - Ensure that hierarchy level 0 has the ranks in the correct scatter gather order. This removes the need for a separate sort list and fixes the offset calculation for allgather. - There were several passes over the hierarchy to determine properties of the hierarchy. I eliminated these extra passes and the memory allocation associated with them and calculate the tree properties on the fly. The same DFS recursion also handles the re-order of level 0. All these changes have been verified with MPI_Allreduce, MPI_Reduce, and MPI_Allgather. All functions now pass all IBM/Open MPI, and IMB tests. coll/ml: correct pointer usage for MPI_BOTTOM Since contiguous datatypes are copied via memcpy (bypassing the convertor) we need to adjust for the lb of the datatype. This corrects problems found testing code that uses MPI_BOTTOM (NULL) as the send pointer. Add fallback collectives for allreduce and reduce. cmr=v1.7.5:reviewer=pasha This commit was SVN r30363.
2014-01-22 19:39:19 +04:00
my_ctl_pointer->flags[BCAST_FLAG] = ready_flag;
for( i = 1; i < pow_2_levels; i++) {
/* get my partner for this level */
partner = my_rank^(1<<i);
partner_ctl_pointer =ctl_structs[partner];
/* Block until partner is at this level of recursive-doubling stage */
while(!IS_SG_DATA_READY(partner_ctl_pointer, ready_flag, sequence_number)) {
opal_progress();
}
assert(partner_ctl_pointer->flags[BCAST_FLAG] >= ready_flag);
if (partner_ctl_pointer->offset < my_ctl_pointer->offset) {
global_sg_offset -= length;
local_sg_offset = global_sg_offset;
} else {
local_sg_offset = global_sg_offset + length;
}
BASESMUMA_VERBOSE(10,("Allgather Phase: Get message from process %d, length %d", partner, length));
mca_bcol_basesmuma_portals_get_msg_fragment_no_eq_h(cs,
&my_ctl_pointer->portals_buf_addr,
&partner_ctl_pointer->portals_buf_addr,local_sg_offset,
local_sg_offset, length);
ready_flag++;
basesmuma: do not call MB directly opal does not always define MB. It is recommended that opal_atomic_[rw]mb is called instead. We will need to address the cases where these functions are no-ops on weak-memory ordered cpus. cmr=v1.7.5:ticket=trac:4158 This commit was SVN r30534. The following Trac tickets were found above: Ticket 4158 --> https://svn.open-mpi.org/trac/ompi/ticket/4158
2014-02-03 21:01:57 +04:00
opal_atomic_wmb ();
coll/ml: add support for blocking and non-blocking allreduce, reduce, and allgather. The new collectives provide a signifigant performance increase over tuned for small and medium messages. We are initially setting the priority lower than tuned until this has had some time to soak in the trunk. Please set coll_ml_priority to 90 for MTT runs. Credit for this work goes to Manjunath Gorentla Venkata (ORNL), Pavel Shamis (ORNL), and Nathan Hjelm (LANL). Commit details (for reference): Import ORNL's collectives for MPI_Allreduce, MPI_Reduce, and MPI_Allgather. We need to take the basesmuma header into account when calculating the ptpcoll small message thresholds. Add a define to bcol.h indicating the maximum header size so we can take the header into account while not making ptpcoll dependent on information from basesmuma. This resolves an issue with allreduce where ptpcoll overwrites the header of the next buffer in the basesmuma bank. Fix reduce and make a sequential collective launcher in coll_ml_inlines.h The root calculation for reduce was wrong for any root != 0. There are four possibilities for the root: - The root is not the current process but is in the current hierarchy. In this case the root is the index of the global root as specified in the root vector. - The root is not the current process and is not in the next level of the hierarchy. In this case 0 must be the local root since this process will never communicate with the real root. - The root is not the current process but will be in next level of the hierarchy. In this case the current process must be the root. - I am the root. The root is my index. Tested with IMB which rotates the root on every call to MPI_Reduce. Consider IMB the reproducer for the issue this commit solves. Make the bcast algorithm decision an enumerated variable Resolve various asset failures when destructing coll ml requests. Two issues: - Always reset the request to be invalid before returning it to the free list. This will avoid an asset in ompi_request_t's destructor. OMPI_REQUEST_FINI does this (and also releases the fortran handle index). - Never explicitly construct or destruct the superclass of an opal object. This screws up the class function tables and will cause either an assert failure or a segmentation fault when destructing coll ml requests. Cleanup allgather. I removed the duplicate non-blocking and blocking functions and modeled the cleanup after what I found in allreduce. Also cleaned up the code somewhat. Don't bother copying from the send to the recieve buffer in bcol_basesmuma_allreduce_intra_fanin_fanout if the pointers are the same. The eliminates a warning about memcpy and aliasing and avoids an unnecessary call to memcpy. Alwasy call CHECK_AND_RELEASE on memsync collectives. There was a call to OBJ_RELEASE on the collective communicator but because CHECK_AND_RECYLCE was never called there was not matching call to OBJ_RELEASE. This caused coll ml to leak communicators. Make allreduce use the sequential collective launcher in coll_ml_inlines.h Just launch the next collective in the component progress. I am a little unsure about this patch. There appears to be some sort of race between collectives that causes buffer exhaustion in some cases (IMB Allreduce is a reproducer). Changing progress to only launch the next bcol seems to resolve the issue but might not be the best fix. Note that I see little-no performance penalty for this change. Fix allreduce when there are extra sources. There was an issue with the buffer offset calculation when there are extra sources. In the case of extra sources == 1 the offset was set to buffer_size (just past the header of the next buffer). I adjusted the buffer size to take into accoun the maximum header size (see the earlier commit that added this) and simplified the offset calculation. Make reduce/allreduce non-blocking. This is required for MPI_Comm_idup to work correctly. This has been tested with various layouts using the ibm testsuite and imb and appears to have the same performance as the old blocking version. Fix allgather for non-contiguous layouts and simplify parsing the topology. Some things in this patch: - There were several comments to the effect that level 0 of the hierarchy MUST contain all of the ranks. At least one function made this assumption but it was not true. I changed the sbgp components and the coll ml initization code to enforce this requirement. - Ensure that hierarchy level 0 has the ranks in the correct scatter gather order. This removes the need for a separate sort list and fixes the offset calculation for allgather. - There were several passes over the hierarchy to determine properties of the hierarchy. I eliminated these extra passes and the memory allocation associated with them and calculate the tree properties on the fly. The same DFS recursion also handles the re-order of level 0. All these changes have been verified with MPI_Allreduce, MPI_Reduce, and MPI_Allgather. All functions now pass all IBM/Open MPI, and IMB tests. coll/ml: correct pointer usage for MPI_BOTTOM Since contiguous datatypes are copied via memcpy (bypassing the convertor) we need to adjust for the lb of the datatype. This corrects problems found testing code that uses MPI_BOTTOM (NULL) as the send pointer. Add fallback collectives for allreduce and reduce. cmr=v1.7.5:reviewer=pasha This commit was SVN r30363.
2014-01-22 19:39:19 +04:00
my_ctl_pointer->flags[BCAST_FLAG] = ready_flag;
/* Block until partner is at this level of recursive-doubling stage */
while(!IS_SG_DATA_READY(partner_ctl_pointer, ready_flag, sequence_number)) {
opal_progress();
}
/* double the length */
length *= 2;
}
/* If I am source for non-power 2 children wait for them */
/* If I am secondary root then my partner would be real root
* so no need for exchange of data with the extra partner */
extra_partner = my_rank + pow_2 ;
if ((extra_partner < group_size) && (!secondary_root)) {
volatile mca_bcol_basesmuma_ctl_struct_t *extra_partner_ctl_pointer = NULL;
extra_partner_ctl_pointer = ctl_structs[extra_partner];
/* Block until extra partner has copied data */
while(!IS_SG_DATA_READY(extra_partner_ctl_pointer, ready_flag, sequence_number)) {
opal_progress();
}
}
Release:
/* free the event queue */
rc = PtlEQFree(allgather_eq_h);
if (rc != PTL_OK) {
BASESMUMA_VERBOSE(10,("PtlEQFree() failed: %d )\n",rc));
}
my_ctl_pointer->starting_flag_value++;
input_args->status = FINISHED;
return BCOL_FN_COMPLETE;
}
/*
* static sg_state_t *sg_state = NULL;
*/
int bcol_basesmuma_lmsg_scatter_allgather_portals_nb_bcast(bcol_function_args_t *input_args,
coll_ml_function_t *c_input_args)
{
int i;
mca_bcol_basesmuma_portal_proc_info_t *portals_info;
int dummy_group_size;
int rc = OMPI_SUCCESS;
int buff_idx;
int count=input_args->count;
size_t pack_len = 0, dt_size =0 ;
struct ompi_datatype_t* dtype=input_args->dtype;
int completed_posts = 0;
sg_state_t *sg_state = NULL;
mca_bcol_basesmuma_module_t *bcol_module = NULL;
int extra_src_posts = -1,allgather_posts = -1, total_msg_posts = -1;
bcol_module = (mca_bcol_basesmuma_module_t *) c_input_args->bcol_module;
/*
sg_state = (sg_state_t*)bcol_module->sg_state;
*/
sg_state = (sg_state_t*)&(bcol_module->sg_state);
/* Re-entering the algorithm */
switch (sg_state->phase) {
case PROBE:
if (input_args->root_flag) {
/* I became a root for this group */
sg_state->phase = START;
goto Start;
}
goto Probe;
break;
case SCATTER_ROOT_WAIT:
goto Scatter_root_wait;
case SCATTER_EXTRA_ROOT_WAIT:
goto Scatter_extra_root_wait;
case SCATTER_PARENT_WAIT:
goto Scatter_parent_wait;
default:
break;
}
sg_state->phase = INIT;
BASESMUMA_VERBOSE(1,("Im entering portals_nb_bcast Unknown root "));
/* Allocate space for algorithm state */
/*
sg_state = (sg_state_t *) malloc(sizeof(sg_state_t));
bcol_module->sg_state = (void *)sg_state;
assert(NULL != sg_state);
*/
sg_state->secondary_root = false;
sg_state->msg_posted = false;
sg_state->matched = 0;
sg_state->phase = SCATTER;
/* Copy input args to local variables */
sg_state->my_userbuf = (void*)((unsigned char*)input_args->userbuf);
assert(sg_state->my_userbuf != NULL);
sg_state->sequence_number=input_args->sequence_num;
sg_state->cs = &mca_bcol_basesmuma_component;
sg_state->bcol_module = (mca_bcol_basesmuma_module_t *) c_input_args->bcol_module;
/* Should this be buffer index (ML) or control buffer index ? */
buff_idx = input_args->src_desc->buffer_index;
/* Initialize SM group info used for control signaling */
init_sm_group_info(sg_state, buff_idx);
/* calculate the largest power of two that is smaller than
* or equal to the group size
*/
sg_state->pow_2_levels = pow_sm_k(2, sg_state->group_size, &(dummy_group_size));
if( sg_state->group_size < (1 << sg_state->pow_2_levels)) {
sg_state->pow_2_levels--;
}
/* power-of-two group size */
sg_state->pow_2 = 1 << sg_state->pow_2_levels;
/* we will work only on packed data - so compute the length*/
ompi_datatype_type_size(dtype, &dt_size);
sg_state->fragment_size = count*dt_size;
/* Init portals scatter allgather info */
rc = init_sm_portals_sg_info(sg_state);
if (rc != OMPI_SUCCESS) {
goto Release;
}
Start :
Extra :
/*
* My rank > pow2 groupsize
*/
if( sg_state->my_rank >= sg_state->pow_2 ) {
if (input_args->root_flag){
rc = sm_portals_extra_root_scatter(sg_state);
if (rc != OMPI_SUCCESS) {
goto Release;
}
} else {
/*
* Wait for my partner to receive bcast data, and copy from it
*/
int extra_parent_rank;
volatile mca_bcol_basesmuma_ctl_struct_t *extra_parent_ctl_pointer = NULL; /* binomial fanout */
extra_parent_rank = sg_state->my_rank & (sg_state->pow_2-1);
extra_parent_ctl_pointer = sg_state->ctl_structs[extra_parent_rank];
sg_state->ready_flag = sg_state->ready_flag + sg_state->pow_2_levels;
while(!IS_SG_DATA_READY(extra_parent_ctl_pointer, sg_state->ready_flag,
sg_state->sequence_number)) {
opal_progress();
}
mca_bcol_basesmuma_portals_get_msg_fragment(sg_state->cs,
sg_state->read_eq,
&sg_state->my_ctl_pointer->portals_buf_addr,
&extra_parent_ctl_pointer->portals_buf_addr, 0,
0, sg_state->fragment_size);
sg_state->my_ctl_pointer->flag = sg_state->ready_flag;
}
goto Release;
}
if (input_args->root_flag) {
BASESMUMA_VERBOSE(1,("Scatter : Im root (bcol_module %x,ctl_pointer %x) my ready flag %d \n",
sg_state->bcol_module, sg_state->my_ctl_pointer, sg_state->ready_flag));
rc = sm_portals_root_scatter(sg_state);
/* gvm Fix: Redudant
basesmuma: do not call MB directly opal does not always define MB. It is recommended that opal_atomic_[rw]mb is called instead. We will need to address the cases where these functions are no-ops on weak-memory ordered cpus. cmr=v1.7.5:ticket=trac:4158 This commit was SVN r30534. The following Trac tickets were found above: Ticket 4158 --> https://svn.open-mpi.org/trac/ompi/ticket/4158
2014-02-03 21:01:57 +04:00
opal_atomic_wmb ();
coll/ml: add support for blocking and non-blocking allreduce, reduce, and allgather. The new collectives provide a signifigant performance increase over tuned for small and medium messages. We are initially setting the priority lower than tuned until this has had some time to soak in the trunk. Please set coll_ml_priority to 90 for MTT runs. Credit for this work goes to Manjunath Gorentla Venkata (ORNL), Pavel Shamis (ORNL), and Nathan Hjelm (LANL). Commit details (for reference): Import ORNL's collectives for MPI_Allreduce, MPI_Reduce, and MPI_Allgather. We need to take the basesmuma header into account when calculating the ptpcoll small message thresholds. Add a define to bcol.h indicating the maximum header size so we can take the header into account while not making ptpcoll dependent on information from basesmuma. This resolves an issue with allreduce where ptpcoll overwrites the header of the next buffer in the basesmuma bank. Fix reduce and make a sequential collective launcher in coll_ml_inlines.h The root calculation for reduce was wrong for any root != 0. There are four possibilities for the root: - The root is not the current process but is in the current hierarchy. In this case the root is the index of the global root as specified in the root vector. - The root is not the current process and is not in the next level of the hierarchy. In this case 0 must be the local root since this process will never communicate with the real root. - The root is not the current process but will be in next level of the hierarchy. In this case the current process must be the root. - I am the root. The root is my index. Tested with IMB which rotates the root on every call to MPI_Reduce. Consider IMB the reproducer for the issue this commit solves. Make the bcast algorithm decision an enumerated variable Resolve various asset failures when destructing coll ml requests. Two issues: - Always reset the request to be invalid before returning it to the free list. This will avoid an asset in ompi_request_t's destructor. OMPI_REQUEST_FINI does this (and also releases the fortran handle index). - Never explicitly construct or destruct the superclass of an opal object. This screws up the class function tables and will cause either an assert failure or a segmentation fault when destructing coll ml requests. Cleanup allgather. I removed the duplicate non-blocking and blocking functions and modeled the cleanup after what I found in allreduce. Also cleaned up the code somewhat. Don't bother copying from the send to the recieve buffer in bcol_basesmuma_allreduce_intra_fanin_fanout if the pointers are the same. The eliminates a warning about memcpy and aliasing and avoids an unnecessary call to memcpy. Alwasy call CHECK_AND_RELEASE on memsync collectives. There was a call to OBJ_RELEASE on the collective communicator but because CHECK_AND_RECYLCE was never called there was not matching call to OBJ_RELEASE. This caused coll ml to leak communicators. Make allreduce use the sequential collective launcher in coll_ml_inlines.h Just launch the next collective in the component progress. I am a little unsure about this patch. There appears to be some sort of race between collectives that causes buffer exhaustion in some cases (IMB Allreduce is a reproducer). Changing progress to only launch the next bcol seems to resolve the issue but might not be the best fix. Note that I see little-no performance penalty for this change. Fix allreduce when there are extra sources. There was an issue with the buffer offset calculation when there are extra sources. In the case of extra sources == 1 the offset was set to buffer_size (just past the header of the next buffer). I adjusted the buffer size to take into accoun the maximum header size (see the earlier commit that added this) and simplified the offset calculation. Make reduce/allreduce non-blocking. This is required for MPI_Comm_idup to work correctly. This has been tested with various layouts using the ibm testsuite and imb and appears to have the same performance as the old blocking version. Fix allgather for non-contiguous layouts and simplify parsing the topology. Some things in this patch: - There were several comments to the effect that level 0 of the hierarchy MUST contain all of the ranks. At least one function made this assumption but it was not true. I changed the sbgp components and the coll ml initization code to enforce this requirement. - Ensure that hierarchy level 0 has the ranks in the correct scatter gather order. This removes the need for a separate sort list and fixes the offset calculation for allgather. - There were several passes over the hierarchy to determine properties of the hierarchy. I eliminated these extra passes and the memory allocation associated with them and calculate the tree properties on the fly. The same DFS recursion also handles the re-order of level 0. All these changes have been verified with MPI_Allreduce, MPI_Reduce, and MPI_Allgather. All functions now pass all IBM/Open MPI, and IMB tests. coll/ml: correct pointer usage for MPI_BOTTOM Since contiguous datatypes are copied via memcpy (bypassing the convertor) we need to adjust for the lb of the datatype. This corrects problems found testing code that uses MPI_BOTTOM (NULL) as the send pointer. Add fallback collectives for allreduce and reduce. cmr=v1.7.5:reviewer=pasha This commit was SVN r30363.
2014-01-22 19:39:19 +04:00
*/
sg_state->my_ctl_pointer->flag = sg_state->ready_flag;
if (rc != OMPI_SUCCESS) {
goto Release;
}
Scatter_root_wait:
BASESMUMA_VERBOSE(5,("Scatter: Im root waiting for children to complete my flag %d",
sg_state->my_ctl_pointer->flag));
for( i = 0; i < sg_state->cs->num_to_probe && completed_posts < sg_state->my_ctl_pointer->n_sends;
i++) {
completed_posts = wait_for_post_complete_nb(sg_state->my_rank,
sg_state->my_ctl_pointer->n_sends, sg_state->ctl_structs,
sg_state->ready_flag, sg_state->sequence_number);
}
if (completed_posts < sg_state->my_ctl_pointer->n_sends) {
sg_state->phase = SCATTER_ROOT_WAIT;
return BCOL_FN_STARTED;
}
goto Allgather;
}
Scatter:
BASESMUMA_VERBOSE(1,("Scatter : Im non-root probing for data "));
/* compute the list of possible sources */
/*
sg_state->src_list = (int *) malloc(sizeof(int) * (sg_state->pow_2_levels + 1));
*/
assert(MAX_SM_GROUP_SIZE > sg_state->pow_2_levels+1);
for( i = 0; i < sg_state->pow_2_levels; i++) {
sg_state->src_list[i] = sg_state->my_rank ^ (1<<i);
}
/* My source might be process > pow_2 */
if ((sg_state->my_rank + sg_state->pow_2) < sg_state->group_size) {
sg_state->src_list[i] = sg_state->my_rank + sg_state->pow_2;
} else {
sg_state->src_list[i] = -1;
}
BASESMUMA_VERBOSE(1,("Scatter : Ready flag %d Im non-root probing for %d procs %d:%d \n",
sg_state->ready_flag,sg_state->pow_2_levels,sg_state->src_list[0],sg_state->src_list[1]));
Probe:
/* If I am not the root, then poll on possible "senders'" control structs */
/* For portals we block for now */
/* Shared memory iprobe */
/*
SG_LARGE_MSG_NB_PROBE(sg_state->src_list, sg_state->pow_2_levels + 1,
sg_state->src_list_index, sg_state->matched, sg_state->src,
sg_state->ctl_structs,
sg_state->parent_ctl_pointer, sg_state->ready_flag, sg_state->sequence_number);
*/
for( i = 0; i < sg_state->cs->num_to_probe && 0 == sg_state->matched;
i++) {
sg_large_msg_probe(sg_state);
}
if (!sg_state->matched) {
sg_state->phase = PROBE;
return BCOL_FN_STARTED;
}
BASESMUMA_VERBOSE(1,("Scatter : Im non-root match received"));
/* If I am a secondary root */
if ((sg_state->matched) && (sg_state->src == sg_state->pow_2 + sg_state->my_rank)) {
BASESMUMA_VERBOSE(5,("Scatter : Im secondary root \n"));
rc = sm_portals_secondary_root_scatter(sg_state);
if (rc != OMPI_SUCCESS) {
goto Release;
}
Scatter_extra_root_wait:
for( i = 0; i < sg_state->cs->num_to_probe && completed_posts < sg_state->my_ctl_pointer->n_sends;
i++) {
completed_posts = wait_for_post_complete_nb(sg_state->my_rank, sg_state->my_ctl_pointer->n_sends,
sg_state->ctl_structs, sg_state->ready_flag, sg_state->sequence_number);
}
if (completed_posts < sg_state->my_ctl_pointer->n_sends) {
sg_state->phase = SCATTER_EXTRA_ROOT_WAIT;
return BCOL_FN_STARTED;
}
goto Allgather;
}
/* we need to see whether this is really
* who we are looking for
*/
for( i = 0; i < sg_state->parent_ctl_pointer->n_sends; i++) {
uint64_t local_offset = 0;
uint64_t remote_offset = 0;
BASESMUMA_VERBOSE(5,("%d found it from %d \n",sg_state->my_rank,sg_state->src));
if( sg_state->my_rank == (sg_state->src^(1<<i))) {
sg_state->parent_ctl_pointer = sg_state->ctl_structs[sg_state->src];
/* we found our root within the group ... */
BASESMUMA_VERBOSE(5,("Shared memory probe was matched, the root is %d ",sg_state->src));
sg_state->my_ctl_pointer->n_sends = i;
/* Am I source for other process during scatter phase */
if ( i > 0) {
BASESMUMA_VERBOSE(1,("Scatter : Im Internal node \n"));
rc = sm_portals_internode_scatter(sg_state);
if (rc != OMPI_SUCCESS) {
goto Release;
}
Scatter_parent_wait:
for( i = 0; i < sg_state->cs->num_to_probe && completed_posts < sg_state->my_ctl_pointer->n_sends;
i++) {
completed_posts = wait_for_post_complete_nb(sg_state->my_rank,
sg_state->my_ctl_pointer->n_sends,
sg_state->ctl_structs,
sg_state->ready_flag, sg_state->sequence_number);
}
if (completed_posts < sg_state->my_ctl_pointer->n_sends) {
sg_state->phase = SCATTER_PARENT_WAIT;
return BCOL_FN_STARTED;
}
} else {
BASESMUMA_VERBOSE(1,("Scatter : Im leaf node \n"));
/* takes care of first level recurssive double */
sg_state->length = sg_state->parent_ctl_pointer->length/
(1<<(sg_state->parent_ctl_pointer->n_sends - 1));
sg_state->my_ctl_pointer->length = sg_state->length;
sg_state->my_ctl_pointer->offset = sg_state->parent_ctl_pointer->offset;
while(!IS_SG_DATA_READY(sg_state->parent_ctl_pointer,
sg_state->ready_flag, sg_state->sequence_number)) {
opal_progress();
}
mca_bcol_basesmuma_portals_get_msg_fragment(sg_state->cs,
sg_state->read_eq,
&sg_state->my_ctl_pointer->portals_buf_addr,
&sg_state->parent_ctl_pointer->portals_buf_addr,
sg_state->my_ctl_pointer->offset,
sg_state->my_ctl_pointer->offset, sg_state->length);
/* signal that I am done reading data from parent */
/*
basesmuma: do not call MB directly opal does not always define MB. It is recommended that opal_atomic_[rw]mb is called instead. We will need to address the cases where these functions are no-ops on weak-memory ordered cpus. cmr=v1.7.5:ticket=trac:4158 This commit was SVN r30534. The following Trac tickets were found above: Ticket 4158 --> https://svn.open-mpi.org/trac/ompi/ticket/4158
2014-02-03 21:01:57 +04:00
opal_atomic_wmb ();
coll/ml: add support for blocking and non-blocking allreduce, reduce, and allgather. The new collectives provide a signifigant performance increase over tuned for small and medium messages. We are initially setting the priority lower than tuned until this has had some time to soak in the trunk. Please set coll_ml_priority to 90 for MTT runs. Credit for this work goes to Manjunath Gorentla Venkata (ORNL), Pavel Shamis (ORNL), and Nathan Hjelm (LANL). Commit details (for reference): Import ORNL's collectives for MPI_Allreduce, MPI_Reduce, and MPI_Allgather. We need to take the basesmuma header into account when calculating the ptpcoll small message thresholds. Add a define to bcol.h indicating the maximum header size so we can take the header into account while not making ptpcoll dependent on information from basesmuma. This resolves an issue with allreduce where ptpcoll overwrites the header of the next buffer in the basesmuma bank. Fix reduce and make a sequential collective launcher in coll_ml_inlines.h The root calculation for reduce was wrong for any root != 0. There are four possibilities for the root: - The root is not the current process but is in the current hierarchy. In this case the root is the index of the global root as specified in the root vector. - The root is not the current process and is not in the next level of the hierarchy. In this case 0 must be the local root since this process will never communicate with the real root. - The root is not the current process but will be in next level of the hierarchy. In this case the current process must be the root. - I am the root. The root is my index. Tested with IMB which rotates the root on every call to MPI_Reduce. Consider IMB the reproducer for the issue this commit solves. Make the bcast algorithm decision an enumerated variable Resolve various asset failures when destructing coll ml requests. Two issues: - Always reset the request to be invalid before returning it to the free list. This will avoid an asset in ompi_request_t's destructor. OMPI_REQUEST_FINI does this (and also releases the fortran handle index). - Never explicitly construct or destruct the superclass of an opal object. This screws up the class function tables and will cause either an assert failure or a segmentation fault when destructing coll ml requests. Cleanup allgather. I removed the duplicate non-blocking and blocking functions and modeled the cleanup after what I found in allreduce. Also cleaned up the code somewhat. Don't bother copying from the send to the recieve buffer in bcol_basesmuma_allreduce_intra_fanin_fanout if the pointers are the same. The eliminates a warning about memcpy and aliasing and avoids an unnecessary call to memcpy. Alwasy call CHECK_AND_RELEASE on memsync collectives. There was a call to OBJ_RELEASE on the collective communicator but because CHECK_AND_RECYLCE was never called there was not matching call to OBJ_RELEASE. This caused coll ml to leak communicators. Make allreduce use the sequential collective launcher in coll_ml_inlines.h Just launch the next collective in the component progress. I am a little unsure about this patch. There appears to be some sort of race between collectives that causes buffer exhaustion in some cases (IMB Allreduce is a reproducer). Changing progress to only launch the next bcol seems to resolve the issue but might not be the best fix. Note that I see little-no performance penalty for this change. Fix allreduce when there are extra sources. There was an issue with the buffer offset calculation when there are extra sources. In the case of extra sources == 1 the offset was set to buffer_size (just past the header of the next buffer). I adjusted the buffer size to take into accoun the maximum header size (see the earlier commit that added this) and simplified the offset calculation. Make reduce/allreduce non-blocking. This is required for MPI_Comm_idup to work correctly. This has been tested with various layouts using the ibm testsuite and imb and appears to have the same performance as the old blocking version. Fix allgather for non-contiguous layouts and simplify parsing the topology. Some things in this patch: - There were several comments to the effect that level 0 of the hierarchy MUST contain all of the ranks. At least one function made this assumption but it was not true. I changed the sbgp components and the coll ml initization code to enforce this requirement. - Ensure that hierarchy level 0 has the ranks in the correct scatter gather order. This removes the need for a separate sort list and fixes the offset calculation for allgather. - There were several passes over the hierarchy to determine properties of the hierarchy. I eliminated these extra passes and the memory allocation associated with them and calculate the tree properties on the fly. The same DFS recursion also handles the re-order of level 0. All these changes have been verified with MPI_Allreduce, MPI_Reduce, and MPI_Allgather. All functions now pass all IBM/Open MPI, and IMB tests. coll/ml: correct pointer usage for MPI_BOTTOM Since contiguous datatypes are copied via memcpy (bypassing the convertor) we need to adjust for the lb of the datatype. This corrects problems found testing code that uses MPI_BOTTOM (NULL) as the send pointer. Add fallback collectives for allreduce and reduce. cmr=v1.7.5:reviewer=pasha This commit was SVN r30363.
2014-01-22 19:39:19 +04:00
*/
sg_state->my_ctl_pointer->flag = sg_state->ready_flag;
}
BASESMUMA_VERBOSE(1,("Completed %d found it from %d \n",
sg_state->my_rank, sg_state->src));
while(sg_state->ready_flag > sg_state->parent_ctl_pointer->flag);
goto Allgather;
}
}
{
/* this is not who we are looking for,
* mark as false positive so we don't
* poll here again
*/
sg_state->src_list[sg_state->src_list_index] = -1;
sg_state->matched = 0;
goto Probe;
}
Allgather:
BASESMUMA_VERBOSE(5,("Completed Scatter phase"));
/* zip it back up - we have already taken care of first level */
sg_state->global_sg_offset = sg_state->my_ctl_pointer->offset;
/* first level of zip up */
sg_state->length = 2 * sg_state->fragment_size/sg_state->pow_2;
/* Posting for all phases of recursive doubling */
extra_src_posts = (sg_state->my_rank + sg_state->pow_2 < sg_state->group_size ) ? 1: 0;
allgather_posts = sg_state->pow_2_levels - 1;
total_msg_posts = allgather_posts + extra_src_posts ;
if ((!sg_state->msg_posted) && (total_msg_posts > 0)){
mca_bcol_basesmuma_portals_post_msg(sg_state->cs, &sg_state->my_ctl_pointer->portals_buf_addr,
sg_state->my_userbuf, sg_state->my_ctl_pointer->portals_buf_addr.userbuf_length,
PTL_EQ_NONE, total_msg_posts, blocked_post,
PTL_MD_EVENT_START_DISABLE| PTL_MD_EVENT_END_DISABLE
| PTL_MD_OP_GET | PTL_MD_MANAGE_REMOTE | PTL_MD_TRUNCATE | PTL_MD_EVENT_AUTO_UNLINK_ENABLE
);
sg_state->msg_posted = true;
}
BASESMUMA_VERBOSE(5,("Done with allgather phase"));
/* I reached an allgather phase */
sg_state->ready_flag++;
basesmuma: do not call MB directly opal does not always define MB. It is recommended that opal_atomic_[rw]mb is called instead. We will need to address the cases where these functions are no-ops on weak-memory ordered cpus. cmr=v1.7.5:ticket=trac:4158 This commit was SVN r30534. The following Trac tickets were found above: Ticket 4158 --> https://svn.open-mpi.org/trac/ompi/ticket/4158
2014-02-03 21:01:57 +04:00
opal_atomic_wmb ();
coll/ml: add support for blocking and non-blocking allreduce, reduce, and allgather. The new collectives provide a signifigant performance increase over tuned for small and medium messages. We are initially setting the priority lower than tuned until this has had some time to soak in the trunk. Please set coll_ml_priority to 90 for MTT runs. Credit for this work goes to Manjunath Gorentla Venkata (ORNL), Pavel Shamis (ORNL), and Nathan Hjelm (LANL). Commit details (for reference): Import ORNL's collectives for MPI_Allreduce, MPI_Reduce, and MPI_Allgather. We need to take the basesmuma header into account when calculating the ptpcoll small message thresholds. Add a define to bcol.h indicating the maximum header size so we can take the header into account while not making ptpcoll dependent on information from basesmuma. This resolves an issue with allreduce where ptpcoll overwrites the header of the next buffer in the basesmuma bank. Fix reduce and make a sequential collective launcher in coll_ml_inlines.h The root calculation for reduce was wrong for any root != 0. There are four possibilities for the root: - The root is not the current process but is in the current hierarchy. In this case the root is the index of the global root as specified in the root vector. - The root is not the current process and is not in the next level of the hierarchy. In this case 0 must be the local root since this process will never communicate with the real root. - The root is not the current process but will be in next level of the hierarchy. In this case the current process must be the root. - I am the root. The root is my index. Tested with IMB which rotates the root on every call to MPI_Reduce. Consider IMB the reproducer for the issue this commit solves. Make the bcast algorithm decision an enumerated variable Resolve various asset failures when destructing coll ml requests. Two issues: - Always reset the request to be invalid before returning it to the free list. This will avoid an asset in ompi_request_t's destructor. OMPI_REQUEST_FINI does this (and also releases the fortran handle index). - Never explicitly construct or destruct the superclass of an opal object. This screws up the class function tables and will cause either an assert failure or a segmentation fault when destructing coll ml requests. Cleanup allgather. I removed the duplicate non-blocking and blocking functions and modeled the cleanup after what I found in allreduce. Also cleaned up the code somewhat. Don't bother copying from the send to the recieve buffer in bcol_basesmuma_allreduce_intra_fanin_fanout if the pointers are the same. The eliminates a warning about memcpy and aliasing and avoids an unnecessary call to memcpy. Alwasy call CHECK_AND_RELEASE on memsync collectives. There was a call to OBJ_RELEASE on the collective communicator but because CHECK_AND_RECYLCE was never called there was not matching call to OBJ_RELEASE. This caused coll ml to leak communicators. Make allreduce use the sequential collective launcher in coll_ml_inlines.h Just launch the next collective in the component progress. I am a little unsure about this patch. There appears to be some sort of race between collectives that causes buffer exhaustion in some cases (IMB Allreduce is a reproducer). Changing progress to only launch the next bcol seems to resolve the issue but might not be the best fix. Note that I see little-no performance penalty for this change. Fix allreduce when there are extra sources. There was an issue with the buffer offset calculation when there are extra sources. In the case of extra sources == 1 the offset was set to buffer_size (just past the header of the next buffer). I adjusted the buffer size to take into accoun the maximum header size (see the earlier commit that added this) and simplified the offset calculation. Make reduce/allreduce non-blocking. This is required for MPI_Comm_idup to work correctly. This has been tested with various layouts using the ibm testsuite and imb and appears to have the same performance as the old blocking version. Fix allgather for non-contiguous layouts and simplify parsing the topology. Some things in this patch: - There were several comments to the effect that level 0 of the hierarchy MUST contain all of the ranks. At least one function made this assumption but it was not true. I changed the sbgp components and the coll ml initization code to enforce this requirement. - Ensure that hierarchy level 0 has the ranks in the correct scatter gather order. This removes the need for a separate sort list and fixes the offset calculation for allgather. - There were several passes over the hierarchy to determine properties of the hierarchy. I eliminated these extra passes and the memory allocation associated with them and calculate the tree properties on the fly. The same DFS recursion also handles the re-order of level 0. All these changes have been verified with MPI_Allreduce, MPI_Reduce, and MPI_Allgather. All functions now pass all IBM/Open MPI, and IMB tests. coll/ml: correct pointer usage for MPI_BOTTOM Since contiguous datatypes are copied via memcpy (bypassing the convertor) we need to adjust for the lb of the datatype. This corrects problems found testing code that uses MPI_BOTTOM (NULL) as the send pointer. Add fallback collectives for allreduce and reduce. cmr=v1.7.5:reviewer=pasha This commit was SVN r30363.
2014-01-22 19:39:19 +04:00
sg_state->my_ctl_pointer->flag = sg_state->ready_flag;
rc = sm_portals_bcasts_allgather_phase(sg_state);
if (rc != OMPI_SUCCESS) {
BASESMUMA_VERBOSE(10,("Error in Bcast's allgather phase "));
goto Release;
}
/* If I am source for non-power 2 children wait for them */
/* If I am secondary root then my partner would be real root
* so no need for exchange of data with the extra partner */
sg_state->extra_partner = sg_state->my_rank + sg_state->pow_2 ;
if ((sg_state->extra_partner < sg_state->group_size) && (!sg_state->secondary_root)) {
sg_state->extra_partner_ctl_pointer = sg_state->ctl_structs[sg_state->extra_partner];
/* Block until extra partner has copied data */
while(!IS_SG_DATA_READY(sg_state->extra_partner_ctl_pointer,
sg_state->ready_flag, sg_state->sequence_number)) {
opal_progress();
}
}
Release:
BASESMUMA_VERBOSE(1,("Im done "));
sg_state->my_ctl_pointer->starting_flag_value++;
sg_state->phase = FINISHED;
return BCOL_FN_COMPLETE;
}
int bcol_basesmuma_lmsg_scatter_allgather_portals_nb_knownroot_bcast(bcol_function_args_t *input_args,
coll_ml_function_t *c_input_args)
{
int i;
mca_bcol_basesmuma_portal_proc_info_t *portals_info;
int dummy_group_size;
int rc = OMPI_SUCCESS;
int buff_idx;
int count=input_args->count;
size_t pack_len = 0, dt_size =0 ;
struct ompi_datatype_t* dtype=input_args->dtype;
int completed_posts = 0;
sg_state_t *sg_state = NULL;
mca_bcol_basesmuma_module_t *bcol_module=NULL;
int extra_src_posts = -1,allgather_posts = -1, total_msg_posts = -1;
bcol_module = (mca_bcol_basesmuma_module_t *) c_input_args->bcol_module;
sg_state = (sg_state_t*)(&bcol_module->sg_state);
BASESMUMA_VERBOSE(1,("Im entering nb_knownroot_bcast bcol = %x ",
c_input_args->bcol_module));
/* Re-entering the algorithm */
switch (sg_state->phase) {
case PROBE:
if (input_args->root_flag) {
/* I became a root for this group */
sg_state->phase = START;
goto Start;
}
goto Probe;
break;
case SCATTER_ROOT_WAIT:
goto Scatter_root_wait;
case SCATTER_EXTRA_ROOT_WAIT:
goto Scatter_extra_root_wait;
case SCATTER_PARENT_WAIT:
goto Scatter_parent_wait;
default:
break;
}
/* Allocate space for algorithm state */
/*
sg_state = (sg_state_t *) malloc(sizeof(sg_state_t));
bcol_module->sg_state = (void*) sg_state;
*/
/* Make sure there userbuffer is not null */
sg_state->phase = INIT;
sg_state->secondary_root = false;
sg_state->msg_posted = false;
sg_state->matched = 0;
/* Copy input args to local variables */
sg_state->my_userbuf = (void*)((unsigned char*)input_args->userbuf);
assert(sg_state->my_userbuf != NULL);
sg_state->sequence_number=input_args->sequence_num;
sg_state->cs = &mca_bcol_basesmuma_component;
sg_state->bcol_module = bcol_module;
buff_idx = input_args->src_desc->buffer_index;
/* Initialize SM group info used for control signaling */
init_sm_group_info(sg_state, buff_idx);
/* calculate the largest power of two that is smaller than
* or equal to the group size
*/
sg_state->pow_2_levels = pow_sm_k(2, sg_state->group_size, &(dummy_group_size));
if( sg_state->group_size < (1 << sg_state->pow_2_levels)) {
sg_state->pow_2_levels--;
}
/* power-of-two group size */
sg_state->pow_2 = 1 << sg_state->pow_2_levels;
/* we will work only on packed data - so compute the length*/
ompi_datatype_type_size(dtype, &dt_size);
sg_state->fragment_size = count*dt_size;
/* Init portals scatter allgather info */
rc = init_sm_portals_sg_info(sg_state);
if (rc != OMPI_SUCCESS) {
goto Release;
}
Start:
Extra :
/*
* My rank > pow2 groupsize
*/
if( sg_state->my_rank >= sg_state->pow_2 ) {
if (input_args->root_flag){
rc = sm_portals_extra_root_scatter(sg_state);
if (rc != OMPI_SUCCESS) {
goto Release;
}
} else {
/*
* Wait for my partner to receive bcast data, and copy from it
*/
int extra_parent_rank;
volatile mca_bcol_basesmuma_ctl_struct_t *extra_parent_ctl_pointer = NULL; /* binomial fanout */
extra_parent_rank = sg_state->my_rank & (sg_state->pow_2-1);
extra_parent_ctl_pointer = sg_state->ctl_structs[extra_parent_rank];
sg_state->ready_flag = sg_state->ready_flag + sg_state->pow_2_levels;
while(!IS_SG_DATA_READY(extra_parent_ctl_pointer, sg_state->ready_flag,
sg_state->sequence_number)) {
opal_progress();
}
mca_bcol_basesmuma_portals_get_msg_fragment(sg_state->cs,
sg_state->read_eq,
&sg_state->my_ctl_pointer->portals_buf_addr,
&extra_parent_ctl_pointer->portals_buf_addr, 0,
0, sg_state->fragment_size);
sg_state->my_ctl_pointer->flag = sg_state->ready_flag;
}
goto Release;
}
if (input_args->root_flag) {
BASESMUMA_VERBOSE(1,("Scatter : Im root (bcol_module %x,ctl_pointer %x) my ready flag %d \n",
bcol_module, sg_state->my_ctl_pointer, sg_state->ready_flag));
rc = sm_portals_root_scatter(sg_state);
sg_state->my_ctl_pointer->flag = sg_state->ready_flag;
if (rc != OMPI_SUCCESS) {
goto Release;
}
Scatter_root_wait:
BASESMUMA_VERBOSE(5,("Scatter: Im root waiting for children to complete my flag %d",
sg_state->my_ctl_pointer->flag));
for( i = 0; i < sg_state->cs->num_to_probe && completed_posts < sg_state->my_ctl_pointer->n_sends;
i++) {
completed_posts = wait_for_post_complete_nb(sg_state->my_rank,
sg_state->my_ctl_pointer->n_sends, sg_state->ctl_structs,
sg_state->ready_flag, sg_state->sequence_number);
}
if (completed_posts < sg_state->my_ctl_pointer->n_sends) {
sg_state->phase = SCATTER_ROOT_WAIT;
return BCOL_FN_STARTED;
}
goto Allgather;
}
Probe:
sg_state->src = compute_src_from_root(input_args->root_route->rank, sg_state->my_rank,
sg_state->pow_2, sg_state->group_size);
sg_state->parent_ctl_pointer = sg_state->ctl_structs[sg_state->src];
while(!IS_SG_DATA_READY(sg_state->parent_ctl_pointer, sg_state->ready_flag,
sg_state->sequence_number)) {
opal_progress();
}
sg_state->matched = true;
/* If I am a secondary root */
if ((sg_state->matched) && (sg_state->src == sg_state->pow_2 + sg_state->my_rank)) {
rc = sm_portals_secondary_root_scatter(sg_state);
if (rc != OMPI_SUCCESS) {
goto Release;
}
Scatter_extra_root_wait:
for( i = 0; i < sg_state->cs->num_to_probe && completed_posts < sg_state->my_ctl_pointer->n_sends;
i++) {
completed_posts = wait_for_post_complete_nb(sg_state->my_rank, sg_state->my_ctl_pointer->n_sends,
sg_state->ctl_structs, sg_state->ready_flag, sg_state->sequence_number);
}
if (completed_posts < sg_state->my_ctl_pointer->n_sends) {
sg_state->phase = SCATTER_EXTRA_ROOT_WAIT;
return BCOL_FN_STARTED;
}
goto Allgather;
}
/* we need to see whether this is really
* who we are looking for
*/
for( i = 0; i < sg_state->parent_ctl_pointer->n_sends; i++) {
uint64_t local_offset = 0;
uint64_t remote_offset = 0;
BASESMUMA_VERBOSE(5,("%d found it from %d \n",sg_state->my_rank,sg_state->src));
if( sg_state->my_rank == (sg_state->src^(1<<i))) {
sg_state->parent_ctl_pointer = sg_state->ctl_structs[sg_state->src];
/* we found our root within the group ... */
BASESMUMA_VERBOSE(10,("Shared memory probe was matched, the root is %d ",sg_state->src));
sg_state->my_ctl_pointer->n_sends = i;
/* Am I source for other process during scatter phase */
if ( i > 0) {
rc = sm_portals_internode_scatter(sg_state);
if (rc != OMPI_SUCCESS) {
goto Release;
}
Scatter_parent_wait:
for( i = 0; i < sg_state->cs->num_to_probe && completed_posts < sg_state->my_ctl_pointer->n_sends;
i++) {
completed_posts = wait_for_post_complete_nb(sg_state->my_rank,
sg_state->my_ctl_pointer->n_sends,
sg_state->ctl_structs,
sg_state->ready_flag, sg_state->sequence_number);
}
if (completed_posts < sg_state->my_ctl_pointer->n_sends) {
sg_state->phase = SCATTER_PARENT_WAIT;
return BCOL_FN_STARTED;
}
} else {
/* takes care of first level recursive double */
sg_state->length = sg_state->parent_ctl_pointer->length/
(1<<(sg_state->parent_ctl_pointer->n_sends - 1));
sg_state->my_ctl_pointer->length = sg_state->length;
sg_state->my_ctl_pointer->offset = sg_state->parent_ctl_pointer->offset;
while(!IS_SG_DATA_READY(sg_state->parent_ctl_pointer,
sg_state->ready_flag, sg_state->sequence_number)) {
opal_progress();
}
mca_bcol_basesmuma_portals_get_msg_fragment(sg_state->cs,
sg_state->read_eq,
&sg_state->my_ctl_pointer->portals_buf_addr,
&sg_state->parent_ctl_pointer->portals_buf_addr,
sg_state->my_ctl_pointer->offset,
sg_state->my_ctl_pointer->offset, sg_state->length);
/* signal that I am done reading data from parent */
sg_state->my_ctl_pointer->flag = sg_state->ready_flag;
}
BASESMUMA_VERBOSE(5,("Completed %d found it from %d \n",
sg_state->my_rank, sg_state->src));
while(sg_state->ready_flag > sg_state->parent_ctl_pointer->flag);
goto Allgather;
}
}
{
/* this is not who we are looking for,
* mark as false positive so we don't
* poll here again
*/
sg_state->src_list[sg_state->src_list_index] = -1;
sg_state->matched = 0;
goto Probe;
}
Allgather:
/* zip it back up - we have already taken care of first level */
sg_state->global_sg_offset = sg_state->my_ctl_pointer->offset;
/* first level of zip up */
sg_state->length = 2 * sg_state->fragment_size/sg_state->pow_2;
/* Posting for all phases of recursive doubling */
extra_src_posts = (sg_state->my_rank + sg_state->pow_2 < sg_state->group_size ) ? 1: 0;
allgather_posts = sg_state->pow_2_levels - 1;
total_msg_posts = allgather_posts + extra_src_posts ;
if ((!sg_state->msg_posted) && (total_msg_posts > 0)){
mca_bcol_basesmuma_portals_post_msg(sg_state->cs, &sg_state->my_ctl_pointer->portals_buf_addr,
sg_state->my_userbuf, sg_state->my_ctl_pointer->portals_buf_addr.userbuf_length,
PTL_EQ_NONE, total_msg_posts, blocked_post,
PTL_MD_EVENT_START_DISABLE| PTL_MD_EVENT_END_DISABLE
| PTL_MD_OP_GET | PTL_MD_MANAGE_REMOTE | PTL_MD_TRUNCATE | PTL_MD_EVENT_AUTO_UNLINK_ENABLE
);
sg_state->msg_posted = true;
}
sg_state->ready_flag++;
basesmuma: do not call MB directly opal does not always define MB. It is recommended that opal_atomic_[rw]mb is called instead. We will need to address the cases where these functions are no-ops on weak-memory ordered cpus. cmr=v1.7.5:ticket=trac:4158 This commit was SVN r30534. The following Trac tickets were found above: Ticket 4158 --> https://svn.open-mpi.org/trac/ompi/ticket/4158
2014-02-03 21:01:57 +04:00
opal_atomic_wmb ();
coll/ml: add support for blocking and non-blocking allreduce, reduce, and allgather. The new collectives provide a signifigant performance increase over tuned for small and medium messages. We are initially setting the priority lower than tuned until this has had some time to soak in the trunk. Please set coll_ml_priority to 90 for MTT runs. Credit for this work goes to Manjunath Gorentla Venkata (ORNL), Pavel Shamis (ORNL), and Nathan Hjelm (LANL). Commit details (for reference): Import ORNL's collectives for MPI_Allreduce, MPI_Reduce, and MPI_Allgather. We need to take the basesmuma header into account when calculating the ptpcoll small message thresholds. Add a define to bcol.h indicating the maximum header size so we can take the header into account while not making ptpcoll dependent on information from basesmuma. This resolves an issue with allreduce where ptpcoll overwrites the header of the next buffer in the basesmuma bank. Fix reduce and make a sequential collective launcher in coll_ml_inlines.h The root calculation for reduce was wrong for any root != 0. There are four possibilities for the root: - The root is not the current process but is in the current hierarchy. In this case the root is the index of the global root as specified in the root vector. - The root is not the current process and is not in the next level of the hierarchy. In this case 0 must be the local root since this process will never communicate with the real root. - The root is not the current process but will be in next level of the hierarchy. In this case the current process must be the root. - I am the root. The root is my index. Tested with IMB which rotates the root on every call to MPI_Reduce. Consider IMB the reproducer for the issue this commit solves. Make the bcast algorithm decision an enumerated variable Resolve various asset failures when destructing coll ml requests. Two issues: - Always reset the request to be invalid before returning it to the free list. This will avoid an asset in ompi_request_t's destructor. OMPI_REQUEST_FINI does this (and also releases the fortran handle index). - Never explicitly construct or destruct the superclass of an opal object. This screws up the class function tables and will cause either an assert failure or a segmentation fault when destructing coll ml requests. Cleanup allgather. I removed the duplicate non-blocking and blocking functions and modeled the cleanup after what I found in allreduce. Also cleaned up the code somewhat. Don't bother copying from the send to the recieve buffer in bcol_basesmuma_allreduce_intra_fanin_fanout if the pointers are the same. The eliminates a warning about memcpy and aliasing and avoids an unnecessary call to memcpy. Alwasy call CHECK_AND_RELEASE on memsync collectives. There was a call to OBJ_RELEASE on the collective communicator but because CHECK_AND_RECYLCE was never called there was not matching call to OBJ_RELEASE. This caused coll ml to leak communicators. Make allreduce use the sequential collective launcher in coll_ml_inlines.h Just launch the next collective in the component progress. I am a little unsure about this patch. There appears to be some sort of race between collectives that causes buffer exhaustion in some cases (IMB Allreduce is a reproducer). Changing progress to only launch the next bcol seems to resolve the issue but might not be the best fix. Note that I see little-no performance penalty for this change. Fix allreduce when there are extra sources. There was an issue with the buffer offset calculation when there are extra sources. In the case of extra sources == 1 the offset was set to buffer_size (just past the header of the next buffer). I adjusted the buffer size to take into accoun the maximum header size (see the earlier commit that added this) and simplified the offset calculation. Make reduce/allreduce non-blocking. This is required for MPI_Comm_idup to work correctly. This has been tested with various layouts using the ibm testsuite and imb and appears to have the same performance as the old blocking version. Fix allgather for non-contiguous layouts and simplify parsing the topology. Some things in this patch: - There were several comments to the effect that level 0 of the hierarchy MUST contain all of the ranks. At least one function made this assumption but it was not true. I changed the sbgp components and the coll ml initization code to enforce this requirement. - Ensure that hierarchy level 0 has the ranks in the correct scatter gather order. This removes the need for a separate sort list and fixes the offset calculation for allgather. - There were several passes over the hierarchy to determine properties of the hierarchy. I eliminated these extra passes and the memory allocation associated with them and calculate the tree properties on the fly. The same DFS recursion also handles the re-order of level 0. All these changes have been verified with MPI_Allreduce, MPI_Reduce, and MPI_Allgather. All functions now pass all IBM/Open MPI, and IMB tests. coll/ml: correct pointer usage for MPI_BOTTOM Since contiguous datatypes are copied via memcpy (bypassing the convertor) we need to adjust for the lb of the datatype. This corrects problems found testing code that uses MPI_BOTTOM (NULL) as the send pointer. Add fallback collectives for allreduce and reduce. cmr=v1.7.5:reviewer=pasha This commit was SVN r30363.
2014-01-22 19:39:19 +04:00
sg_state->my_ctl_pointer->flag = sg_state->ready_flag;
rc = sm_portals_bcasts_allgather_phase(sg_state);
if (rc != OMPI_SUCCESS) {
BASESMUMA_VERBOSE(10,("Error in Bcast's allgather phase "));
goto Release;
}
/* If I am source for non-power 2 children wait for them */
/* If I am secondary root then my partner would be real root
* so no need for exchange of data with the extra partner */
sg_state->extra_partner = sg_state->my_rank + sg_state->pow_2 ;
if ((sg_state->extra_partner < sg_state->group_size) && (!sg_state->secondary_root)) {
sg_state->extra_partner_ctl_pointer = sg_state->ctl_structs[sg_state->extra_partner];
/* Block until extra partner has copied data */
while(!IS_SG_DATA_READY(sg_state->extra_partner_ctl_pointer,
sg_state->ready_flag, sg_state->sequence_number)) {
opal_progress();
}
}
Release:
BASESMUMA_VERBOSE(1,("Im done "));
sg_state->my_ctl_pointer->starting_flag_value++;
sg_state->phase = FINISHED;
return BCOL_FN_COMPLETE;
}
#endif /* __PORTALS_AVAIL__ */
Копировать постоянную ссылку